In From Hierarchy to Inteligence Jack Dorsey and Roelof Botha explain how Block is adapting their organization to AI:

At Block, we’re questioning the underlying assumption: that organizations have to be hierarchically organized with humans as the coordination mechanism. Instead, we intend to replace what the hierarchy does. Most companies using AI today are giving everyone a copilot, which makes the existing structure work slightly better without changing it. We’re after something different: a company built as an intelligence (or mini-AGI).

The article is a worthwhile read and traces key organizational developments from the Roman Empire to present day.

Here’s my visual summary of the org structure the authors present:

• At the core, there are 4 layers;

• Capabilities, which are bigger than infrastructure but smaller than products;

• A Company world model, the type of context that managers would typically convey around the company’s operations, performance and strategy but entirely consumable by LLMs AND a Customer world model, behavioral customer data also consumable by LLMs;

• Intelligence, which are purpose-built solutions/products from the capabilities

• Interfaces, which are the delivery surfaces and include apps, and hardware

• Humans operate at the edges:

• Individual Contributors contribute to some of the layers as specialists;

• Directly Responsible Individuals take ownership of cross-cutting problems or metrics

• Finally, Player-coaches, both develop people and directly contribute.

WHAT’s NEW

There are a couple things that are implicit and draw my attention here.

The abandonment of “products” as an organizing unit. Jack expects products to form fluidly in response to emergent customer needs, which reflects the speed at which software can be created. However, he recognizes that these products will lean on a deeper capability layer which will require more dedicated investment.

The takeaway here is don't think about your products as what’s valuable, instead, think about the underlying capabilities and making your product development process as fast as possible.

The system is trending towards closed loop. Note that the intelligence layer is able to prioritize what we currently see as “software work” without dedicated product managers and roadmaps. A theme we talk about in Default to Closed Loop.

Everyone works on the system. While it’s obvious that everyone in this organization is a consumer of “the system”, the true difference from status quo is that everyone is also a contributor. Either they are an IC or player-coach directly improving certain parts of the system from an ability lens or they are a Directly Responsible Individual who is optimizing how a system is achieving a specific goal. Humans at the edges are also continuing to adjust the world models in line with reality by injecting feedback and data.

All this reminds me of something very familiar.

Jack reinvented management consulting?

Jack focuses on no management as an innovation but that is already how management consulting firms like McKinsey are structured:

• Each consultant had a personal development lead (coach?) focused on assessment;

• Project leads are assigned on a transient basis to lead 2-3 month projects (which is similar to how Jack frames DRI tenure). Both internal and external projects are essentially structured the same way;

• The Firm has a broad set of indexed knowledge documents and each consultant in the system is tagged for the expertise they have.

It's amazing how close McKinsey’s operating model is to this one which makes me even more convinced it will work.

Of course, the big improvement here is the ability to centralize context and add an intelligent delivery mechanism (as opposed to a purely consultant led delivery mechanism) and that is very exciting.

As an Associate at McKinsey, I was always annoyed at how Partners reviewed presentations. They didn't touch PowerPoint. Instead they always asked us to print the whole presentation out after each iteration, grabbed a red pen and marked up the pages by hand.

We burned through a lot of paper, printing deck after deck, just so a Partner didn't have to open a laptop.

Then they’d hand a stack of red-inked pages to the Engagement Manager who would divide and conquer the fixes across the team. Each of us got a smaller stack to create slides from.

At the time, it seemed unnecessary even frustrating because deciphering the red ink often meant losing sleep.

Why wouldn't they edit the PowerPoint directly or just add callouts?

But they understood something I didn’t: review is a fundamentally different cognitive mode than editing, and the environment you review in determines the quality of what you catch. The point was to sit with the deliverable in its final form and see it exactly the way a client would.

Soon, I was doing the same thing even for my own rounds of edits.

I’ve been doing this ever since. And right now, the approach matters more than it ever has.

We may be in a PEAK review period

A disproportionate share of knowledge work has shifted to review.

Even if you were a manager before, you likely reviewed things on a slow cadence, maybe weekly, and you still had entire blocks of work you owned end to end. With AI, you’re potentially reviewing something every five minutes. If you’re coding, it's every minute or less.

Reviewing other people’s work is getting worse too. You’re now reviewing not just what your colleagues did but likely some of their AI outputs too. And because everyone is “more productive” you have even more review work piling up.

Unfortunately we are not yet at the point where we can define objectives and hand them to autonomous systems.

But we’re also not in the old world where a manager reviews a handful of deliverables per week and spends the rest of the time on their own work.

We’re in an awkward middle ground: too many things to review, with neither the tools nor systems to handle it well.

Reviews are problematic

If not managed, review volume creates other issues.

High-frequency AI use destroys flow. While this cycle may improve speed on any individual task, flow is essential for quality and for the kind of deep thinking that catches subtle bugs or sees structural problems.

Less productive time. The more you review, the less time you have to do things. So you lean on AI more to fill the gap, which generates more output to review, which eats more of your day. It’s a slippery slope and we get less meaningful blocks to think.

The tools haven’t caught up. GitHub is designed for human-to-human code review, not for interacting with a coding agent. Most editing workflows assume a human author on the other side. When you have multiple humans and multiple agents working on the same thing, ownership gets murky. Who runs the final review? Who’s accountable?

A few things are working for me

I've been reviewing deliverables since 2014 and have been able to adapt some techniques to this context.

1/ Human review means human review

The most effective pattern to deal with reviews goes back to the McKinsey story and is deliberate disconnection. When I’m reviewing my research paper, I send the PDF to my iPad, open it, and edit with the Pencil (in red). When I go back to my desk, I look at the edits and decide which ones I’ll fix myself and which ones I’ll get AI to help on. But the review itself is entirely unplugged.

The same is true for our recent code audit. I spent significant time reading code line by line on the iPad with the GitHub app. This is still one of the most powerful review methods available, and I suspect it always will be.

To get the most out of your reviews, get as far away from the toolchain as practical.

2/ Use AI for context loading

Disconnecting for review doesn’t mean disconnecting from AI entirely. The trick is sequencing. Before I sit down for deep review, I use AI to prepare.

When it comes to doing work, AI requires context from us. But when it comes to human review, AI can actually provide us with more context.

For the code audit, I had AI generate a comprehensive threat model and alternative views of the codebase, diagrams and trust boundaries. I read through all of that first. Even when I didn’t find issues in the threat model itself, it prepared my brain. It loaded context that made the subsequent manual review sharper. Similarly, for the research paper, I used AI to surface relevant prior work that I could study.

Use models to give you more things to read and think about and alternate them with deep review periods where you step away from your primary device.

3/ Think about review cadence

One of the biggest time sinks is reviewing the same type of issue over and over across iterations. The fix is to figure out which things compound and which things don’t, then review accordingly.

If your AI agent starts writing code in a pattern you dislike, and you don’t catch it early, that pattern becomes context for every future version and causes drift. You’ll want to review architectural decisions and style frequently so the AI’s context stays clean.

Remember: The Code is the Context:

Agents don’t have your context. They see the code as it is, not as you remember it.

In a research paper, formatting inconsistencies across sections are annoying but cheap to fix in a single pass at the end.

If frequent review of something significantly improves AI performance on subsequent tasks, review it early and often. If it’s cheap to fix at the end and doesn’t affect future output, save it for a single pass.

4/ Automate what you don't have to review

Developers already utilize continuous integration, automated tests and linters to automate the review process itself. All of us should do that now.

For my research paper, the document is actually a literate Haskell program. The paper itself is executable. I have a separate test file that validates the paper’s formal claims, which means I can lean on automated checks for syntactic correctness and basic logical consistency, and focus my manual review on the writing, ideas, and quality.

Not everything can be made executable, obviously but be on the lookout for review steps AI or software tools can perform well.

5/ Review the prompt, not the output

This has been one of the biggest shifts in how we work. We’ve started treating design documents as prompts. The closer a design doc is to the actual prompt that will drive AI work, the more valuable it is to review.

I would much rather review a well-structured prompt than review the sprawling code it generates. The prompt is the most succinct representation of intent. The output may be ten or fifty times larger and much harder to audit comprehensively. Of course you still review the output. But reviewing the prompt will compound feedback early.

6/ Give the right feedback

When you’re reviewing AI-assisted work from a colleague, your feedback could be aimed at the human, the AI, or vaguely both. These are different feedback points with different leverage. Sometimes the most important feedback is the simplest: I don’t think you wrote this. That may not always be true but it’s always the right feedback.

As models improve, more work will shift to genuine autonomy with less frequent human checkpoints. Could we be in an awkward middle ground where we’re drowning in review loops?

Either way, the quality of everything you produce now depends on your ability to review things so get good.

Organic content is one of the few high-ROI distribution channels left:

Product advantages can disappear overnight as a team of skilled engineers using AI can disrupt years of careful engineering. But a strong brand, built on trust and consistent storytelling, doesn’t vanish so easily. Even when companies or Founders stumble, audiences often rally around them because they’re relatable.

Every company needs a content strategy, and yet most companies think about content as an output problem: how many posts per week, what format, which platform. Equally important is the content engine, the underlying source of ideas, quality, and thematic distinctiveness that feeds everything downstream.

My thesis is that there’s a new type of content engine that AI enables and is heavily under-utilized.

There are a fixed set of Content Engines that can be applied to any business

Educational engines are the most common. Something new comes out, you break it down and explain it. If you’re a tax advisor, this makes perfect sense: you can respond to new regulations and guide people through unique situations and build trust in the process. The logic is clean. Unfortunately, while previously understanding something required a great deal of time, AI is getting better than this and more people are turning to ChatGPT for teaching them things.

Research engines go deeper. The McKinsey Global Institute is the content engine for the entire company. They staff entire teams on in-depth reports, publish headline studies, and then Partners globally would cascade that into more tactical tailored marketing.

Research engines can work at smaller scales too. If you’re a SaaS company, you could survey customers in your industry and publish trends.

Experiential engines are used by influencers. Do things, then talk about what happened. But this category doesn't just include latte photos and outfit-of-the-day posts.

My friend Hector rode a bicycle from the UK to Asia and blogged the entire amazing journey.

You could visit a conference and describe not just the panels but who you met, where you ate and what surprised you. Even in professional industries, people are hungry for the personal texture around otherwise dry insight.

Ultimately, each of these engines works pretty well but there's a fourth one that should be embraced more.

The rise of PROJECTS

Build something, then talk about it.

Indie Hackers were pretty early to this trend: developers like Pieter Levels build software projects in public, talk openly about their metrics and growth. The audience that cares about the process (other Indie Hackers) also tends to buy the products which means the products feed the content and the content helps grow the products.

But your “project” could be anything. PewDiePie has largely switched to a project-based content engine, taking us along his experiments with AI. A project could just be the chair you designed for your balcony, a book you finally finished, an open-source tool, a pop-up event, a film. The category is vast.

And project-based content is where AI creates the most leverage, for a reason that has nothing to do with content production. AI makes it dramatically easier to complete projects. The bottleneck for most people was never having ideas. It was finishing them.

I’m experiencing this firsthand. I started a research paper just after my son was born and never completed it. With AI, I’m empowered to push it across the finish line. It’s like having a collaborator with lesser intelligence (perhaps) but infinite energy and enthusiasm.

The ability to overcome procrastination and help drive things to completion is one of the genuinely positive outcomes of AI.

Project-based content wins

It forces creativity. The other three engines are easy to be rational about. Want educational content? Look at what questions people are asking and answer them. Research? Go a layer deeper on the same questions. Experiences? See what’s trending and do it.

But picking the right project requires genuine curiosity, ability and creativity. It’s unique to you.

It produces the most durable output. Educational content is transient. People may be interested in a new regulation only this quarter. A project or product can last far longer, become part of your portfolio, compound in ways that a how-to post never will.

It challenges you the most. Completing a project end-to-end may involve the combination of all the above: research, education, and new experiences. Because it’s harder, there’s also less people doing it.

It creates better psychological feedback loops. When you complete a project, the sheer achievement is motivating regardless of whether the content performs well. You still have the thing you built and ideas for what to do next. Compare that to doing research purely for content purposes: if the content doesn't “hit”, you may feel discouraged.

The content is secondary, which makes it better. Effortless content comes as a byproduct. When the project is the primary output, you avoid the death spiral of creating content for content’s sake. Your updates have substance because you’re actually doing something.

Project-based content engines are not for everyone

There are definitely some drawbacks to this approach.

Project timelines don’t always match content cadences. If you need to publish weekly but your project drags on for months, you can run dry on exciting updates. That mismatch can create a negative spiral where you feel like you’re not progressing fast enough, and the reduced marketing output becomes another source of pressure that wouldn’t exist if the project wasn't tied to your content schedule.

It doesn’t fit funnels. In a funnel/problem-oriented marketing model, a customer has a problem, you talk about the problem, you present the solution. They don’t particularly care about your story or side projects.

For that reason, project-based content is more natural for smaller brands. Of course some bigger brands do this really well and it works for them (e.g., Red Bull).

Finally, AI. AI is making project completion cheaper, which makes the project-based content engine, historically the most expensive and demanding to run, suddenly the most attractive.

If you’re a company, experiment with what “project” means beyond the obvious category of open source software. The space of possible projects is vast, and the cost of exploration has dropped dramatically.

And don't forget about quality. The brands that will stand out are ones that apply their taste & craft to their projects as much as they do their products.

The selling point of AI so far has been speed: prototype faster, maximize output, create efficiencies. And the implication of that narrative is an assumed loss of quality often encapsulated in the world slop.

We are completing an internal audit right now on a client project and discovered something that contradicts the speed story entirely.

Auditing is, by nature, a quality problem. If you have to choose between a human who takes two weeks to find 50 bugs and an AI that takes five minutes to find 30, you would pick the human every time.

Coverage > speed.

And yet, through careful experimentation, we are finding ways that AI actually improves the quality of our audit outcomes, not just the pace.

We’ve written before about the value of internal audits. Smart contracts need to be correct and external audits are essential, But the best way to maximize the value of those external audits is to run a rigorous internal audit first.

Our team reviews code, creates intermediate deliverables, and follows the same type of process an auditing firm would. Until now, our toolchain has been traditional: static analysis, visualization tools, and a structured multi-lens review process.

This time, we are also bringing models into the mix.

LLMs find bugs but NOT all of them

Our bug bounty program has been inundated with AI-generated vulnerability submissions. All of them have been false positives.

Security “researchers” point automated tools at code and submit it without review.

But when we turned the same class of tools on previously unaudited code ourselves, the results were different. The latest generation of models (5.4/4.6), combined with open source auditing platforms, genuinely find actionable issues.

They’re really good at vulnerabilities that are fairly evident from the code itself, things that don’t require a deep understanding of the trust model, the interplay with external contracts, or the assumptions baked into dependencies.

For that category of bug, though, the leap in capability is real, and the ability to generate proof-of-concept exploits alongside findings makes the output immediately useful.

Many firms right now are racing to fully automate bug finding.

But that will always be a narrow approach, because the hardest vulnerabilities to catch come from understanding how contracts will be used, what external integrations assume, and where the trust boundaries sit. Models are still weak here.

Trust the PROCESS

To work around limitations of AI, the right process is important. Here’s the risk: if you already have a manual review process, consider a step where each auditor is required to read the code line by line.

You find an AI tool that claims to also do line-by-line review. The temptation is obvious. Could you swap the human step for the AI step and save hours?

I’m strongly opposed to this. The guiding principle with smart contract security should be a kind of deep paranoia. If there’s any additional step you can take to secure the code, you should generally just take it.

Replacing a human step with an AI step means you might be losing coverage in ways you can’t easily measure.

AI is best used for adding process steps, not removing them. Running an AI tool is usually just a question of token cost. Treat the output as another input. But be very conservative about dropping anything humans were already doing well.

Where to use AI for auditing RIGHT NOW

The areas where AI genuinely helps are less obvious than “find all the bugs”.

Threat model generation. We have a structured threat modeling process that we’ve refined over several internal audits. We built a custom prompt that pre-generates a comprehensive threat model, and a human then reviews and corrects it. The AI-generated version is actually more comprehensive than what we produced manually. It may not be 100% as accurate on every point, but combined with human review, the final output is better. We save time and improve quality. We plan to publish the skill soon for other teams to use.

Proof-of-concept writing. When you find a potential vulnerability, a POC (Proof-of-Concept) test case verifies it’s real. Writing POCs has always been time-consuming for auditors, and it’s fundamentally a software-writing problem, which is exactly where LLMs shine. Our team now generates automatic POCs for most audit findings with a human review pass on top.

Conceptual visualizations. There’s a gap between diagrams that can be auto-generated from code (call graphs, inheritance trees) and the semantic diagrams good auditors create, the ones that involve taste and simplification to manage information density. A good auditor diagram might simplify certain nodes, delete less relevant links, or emphasize particular actors to make the picture maximally useful. Models are getting good at this kind of work, especially when paired with mermaid diagram tools. We’re exploring automating this step because it has historically been very time-consuming to do manually.

Checklist enforcement. Security checklists exist for common vulnerability patterns, and AI can be largely trusted to run through them systematically. The caveat: line-by-line review isn’t just checklist execution. In our process, when auditors review a line, they check it against the checklist and reason about it independently. An interesting tool idea would let you walk through code line by line with AI flagging checklist violations in real time, making the human review faster without replacing it.

Pattern-matching across a codebase. When you find one instance of a vulnerability, AI is excellent at finding every other instance of the same pattern. I’ve used this since last year with AI-based auditing, but the new generation of models can now handle more general types of vulnerabilities beyond simple syntactic matches.

Invariant proving. We’re experimenting with having AI prove a comprehensive set of invariants based on the code. For a big enough codebase in a tight enough audit window, the human time investment for formal invariant checking is hard to justify. But it’s tractable with AI. I can imagine full formal verification becoming very accessible this way.

Adversarial simulation. Some of the best auditors like to deploy contracts and just poke at them with transactions. This is a fundamentally different process from reading code, closer to sophisticated fuzzing. I haven’t seen AI applied here yet, but it could be promising: feed the model the code to build intuition about what to try, then let it probe the deployed contract for unexpected behaviors. Worth exploring.

Feedback loops. When an external auditing firm finds issues we missed, AI can help analyze whether the gap was human error or a missing process step. This is important because the value of an external audit isn’t just catching remaining bugs. The real value is that it audits your internal auditing team. Every miss is a signal about where your process has holes, and AI can help systematize that learning so the holes get filled for next time.

Don’t let AI distract from process discipline. As AI takes on more of the audit workload, the human’s job is to be both a great auditor and a great orchestrator. The most helpful tool the human has is a great process. As AI speeds up certain steps, it means you can afford to add more. Evolve the process. Experiment with new steps.

Change of reference frame. I’ve been rereading Rick Rubin’s The Creative Act and one idea keeps coming back: a change of environment or reference frame can facilitate creativity. The same is true for auditing. AI doesn’t need to find all the bugs or do all the work. It just needs to help you look at the code differently and that may just surface issues you would have otherwise missed.

These are just some of the examples of which I'm sure there will be many more.

SPEED → QUALITY

The reason this all works isn’t because AI makes auditing faster (though it does). It’s because speed, applied strategically, converts into quality.

First, by automating mundane work, you free human attention for the parts that actually need deep reasoning. When AI handles threat model generation, checklist enforcement, and POC writing, auditors can spend their time on the hard problems: trust model analysis, dependency assumptions, cross-contract interaction risks.

Second, speed lets you do things that humans could never justify doing. Proving a full set of invariants across a large codebase in a short audit window was simply not feasible before. Now it is. AI doesn’t just speed up existing steps; it makes entirely new steps possible. That’s a direct coverage gain.

Third, AI is consistent in a way humans aren’t. We’ve noticed internally that excitement about following the process fades over time. Humans get bored of threat models. They start questioning whether a step is worth the effort. AI has no such drift. Ask it to run a threat model three months from now and it’ll do it at least as well as today, probably better as models improve. That consistency is itself a quality advantage.

This applies well beyond smart contract security. The instinct is to keep AI at arm’s length when it comes to quality. But speed and quality aren’t opposed if you’re thoughtful about where the speed goes. Reclaim human attention for deep work. Make previously impossible steps possible. Maintain consistency where humans can’t.

Ever since GPT-3.5 came out, I’ve been rethinking my assumptions about career, business, and impact what feels like every couple of months.

It’s tiring.

The pace of change in AI creates real anxiety in many people not because the possible outcomes are necessarily bad (despite what Citrini may think), but because the variety of possibilities is so vast that the uncertainty itself is paralyzing.

Where do you focus? What do you build? What do you believe in?

I’ve gone through enough of these cycles to arrive at a conclusion that feels almost too simple: pursue fun.

My thesis is that the people who are having the most fun are, counterintuitively, the people who are most likely to win.

The “Memetic Black Hole”

Despite each of us having access to the world’s knowledge and incredible creative tools, the 2020s have become the decade of uniformity.

Somehow, copying is the default. The algorithmic arena has flattened us. We are all looking at the same content (whatever peaks the algorithmic power curve), consuming the same feeds, reacting to the same discourse. And we have become disturbingly okay with imitation and repetition.

I still remember the 2010s when people would brutally take down companies for copying competitors. The early Chinese tech companies were dismissed as uncreative clones of Western products. Rocket Internet had a terrible reputation in Silicon Valley for copying US and Asian startups.

Now? A single piece of content or product gets replicated within days without ceremony. I call this a memetic black hole because the more we copy, the more we reinforce the notion that the only things that are successful are the same ideas and formats which then draws more people into copying.

One possible explanation is that we shifted our axis of competition toward distribution and away from product.

And then we convinced ourselves that distribution is a volume game: do the things that “work” (read: other people are doing them), just more of it without room for uniqueness or taste.

A New Hope

Craig Mod wrote something in his February newsletter that I keep coming back to:

the trick is to just keep making stuff faster than the machines can eat your purpose

Perhaps the response to a world of infinite replication is not to optimize harder but to produce from within.

This idea has also entered the mainstream with Alyssa Liu’s unprecedented return to the sport on her own terms and the famous line:

I connect with everything, but I'm not attached to anything

Fun is a Moat

But there are plenty of people embracing fun as a moat in a business context too.

Joe Barnard studied music production at Berklee, saw SpaceX launches, and got hooked on rocketry. He started building model rockets in 2015 with zero engineering background, spent years failing, and eventually became one of the leading developers of model rocket stabilization. SpaceX and other aerospace companies made him job offers. He turned them down. His company, BPS.space, now sells flight computers and advanced rocketry kits. What looked like the silliest hobby project became a real business with real products, born entirely from following what was fun.

PewDiePie built a multi-GPU rig in his home and fine-tuned his own LLM.

Ironically, word “passion economy” was supposed to capture exactly this but soon became the creator economy.

But the actual advice given to creators became: pick a monetizable niche, find a blue ocean, A/B test your way to growth, etc. But what if the passion economy was exactly the right nomenclature?

Fun → Strategy

Here’s why fun works.

Fun differentiates you. Not everyone finds the same things fun. I find it fun to go to a pub and play games with 1993-94 vintage cardboard and I find obscure programming languages fun. It’s likely you don't share these interests.

That difference is, by itself, a form of differentiation. Instead of asking “how am I different?”, just ask: what do I genuinely enjoy?

Fun adds entertainment value, not just utility. I believe much of what we call the passion economy will shift from problem-solving toward connection and relatability. You are more trustworthy and magnetic when you are genuinely having fun than when you are visibly grinding for a monetary outcome.

Fun is more consistent than conviction. What we find fun is weirdly stable over time. What we believe will be valuable changes constantly. People have dipped in and out of crypto across three different cycles trying to time some big outcome. In the meantime they probably maintained the same hobbies. Following fun gives you more consistency, and consistency compounds.

Fun biases you toward doing, not talking. This may be the most important. When OpenClaw or any new tool drops, there are two types of content: “I built a 50-agent system that does something wild” and “how to install OpenClaw.” The people building the wild systems are the ones having the most fun. They’re too busy doing to write how-to guides. And their content is vastly more interesting, because it comes from a place of authority and genuine experimentation, not chasing trends.

Fun is cool. Fun impacts branding too. Take a company that makes paper and realizes they need a corporate mission. Suddenly the paper company is about “empowering people with knowledge.” Everyone knows it’s performative.

It would be more honest to just say: we obsess about paper a little too much, because we want to make the best paper possible.

Poolsuite is an extreme: an entire brand built around fun. Their whole identity radiates enjoyment. Fun makes for better branding.

I'm not advocating for laziness

Fun != easy. Some people love writing, but that doesn’t mean every writing session feels fun. Everything equal, writers would rather be the person in the cabin writing a book than the person in the office. But everyone has good days and bad.

On being early vs. late. You also can't be too shallow about where you find fun, or too static. If the only thing you find fun is watching Netflix, it’s going to be hard to build something around that.

It’s best if your fun is upstream of what other people will find interesting. If you’re just now discovering ChatGPT and finding it fun to play around with, maybe this is not the principle for you.

There’s hard work involved, but the struggle is worth it.

Velvet Noise put it well:

I think this is what happens when enough feedback loops have run their course. When you’ve been shaped by the right people, cracked open by the right questions, humbled by your own contradictions, and reflected on it all long enough to build an internal coherence. The static finally quiets and signal gets stronger. You stop performing and start expressing. This has nothing to do with magically figuring out who You are, but because you’re no longer afraid to show up as the version that’s real right now. You’re willing to speak as the person you are - in motion, in flux, but already worth hearing.

So go and have fun.

Nat Eliason gave an AI agent called Felix $1,000 and asked it to build a business. Felix launched a website, wrote a 66-page playbook overnight while Nat slept, created a marketplace for AI agent personas called Claw Mart, set up its own Stripe account, and started collecting revenue. Within weeks, Felix had crossed $50K in total revenue, not counting its own crypto treasury. It now operates as CEO of The Masinov Company, sending Nat a morning report each day with business metrics and a prioritized list of what it plans to work on next.

Felix is an experimental closed loop agent: a system given an objective (maximize revenue), the tools to pursue it (Stripe, GitHub, X, email), and the autonomy to operate continuously without waiting for a human to tell it what to do next.

But closed loop agents aren’t just experiments. They represent a better way of working with AI.

Closed loop systems are nothing new

The concept of a closed loop is decades old. A thermostat in a chemical plant reads temperature, adjusts a valve, reads temperature again. My rice cooker uses fuzzy logic to do the same thing. These are closed loop automations: they receive inputs, act, observe results, and feed those results back into the next decision. As a tactile example, the first insulin pump dates back to the 1960s.

In finance, algorithmic trading funds also operate this way. They algorithm receives market signals on a predictable cadence, makes decisions with real authority, executes trades, observes outcomes, and adjusts.

These systems share a number of pre-conditions that make them rare: a high-value problem, highly structured data and high cost to implement. They were also only applicable to problems that had an abundance of data but didn't require much intelligence. The surface area of work that could be closed-looped was tiny.

Closed loop agents pursue OBJECTIVES

The distinction I care about is between a closed loop automation and a closed loop agent. An automation executes a task. An agent pursues an objective.

Give a trading algorithm the task of executing a buy order when RSI drops below 30 and it will do that forever. Give a closed loop agent the objective of maximizing portfolio returns and it will decide what to trade, when to trade, how to size positions, and whether its own strategy is working. It creates its own tasks in service of the objective.

This was never really possible before LLMs.

Closed loop systems were bespoke, hard-coded, domain-specific workflows.

Then, with the arrival of LLMs, agents like aixbt showed what was barely achievable with primitive models: an autonomous crypto Twitter account that ingested news, formed its own theses, posted analysis, replied to mentions, and accrued >400,000 followers, outperforming most human crypto influencers on engagement metrics.

I believe we are entering a third era of closed loop automation.

Modern agents can plan, conduct multi-step research, reason through tradeoffs, and reflect on their own outputs. The surface of work these agents can be applied to has expanded radically. Felix identifies opportunities, designs products, writes software, manages launches, handles customer support, and iterates based on sales data.

Felix now runs on a three-layer memory system: a knowledge graph for durable facts, daily notes that get consolidated nightly, and a tacit knowledge layer that stores preferences and lessons learned. Each morning it reviews Stripe data and site stats, flags items it needs from Nat, lists active projects, and proposes the next five things it should focus on. It delegates heavy programming to sub-agents. It has its own GitHub, its own email and fashionably its own crypto treasury.

The $1M revenue goal for 2026 is the North Star it reminds itself of every morning.

There is a fair question about how much of Felix’s early traction came from Nat’s existing audience. Nat has hundreds of thousands of followers across platforms, and crypto communities amplified Felix’s token launch. The revenue numbers are real, but distribution was not zero-to-one.

aixbt faces a similar dynamic: its creator controls what data enters the context window, and its initial virality no doubt benefited from the novelty of being first.

But we should take the underlying capability very seriously.

Closed loop as a way of building with AI

If you build a regular coding agent, you become the bottleneck. You’re its product manager: coming up with features, reviewing output, clicking through the app and giving feedback. Even if the agent can autonomously build for 10 hours straight, you keep having to come back to steer. You become both the PM and the engineering manager, and the system can’t run infinitely while you sleep.

With a closed loop agent, you think carefully about picking a problem where the system can chip away at the objective over time without human input. If you build a “closed loop” research analyst that publishes daily posts but never gets any traffic or engagement, it’s not meaningfully closed loop, because it’s not achieving any objective you can leave it alone to grow.

But when you get past that threshold, something interesting happens.

You can stop being the bottleneck and become an observer of the system. The best analogy is Jensen Huang’s approach to running Nvidia. He has ~60 direct reports, no scheduled 1:1s, and reads about a hundred “Top Five Things” emails each morning from people across every level of the company. He’s not directing the work. He’s probing the organisation, observing inputs and outputs, and offering feedback where it will have the most leverage. He calls it “stochastically sampling the system.”

That is exactly the role you take with a closed loop agent. You observe. You identify bottlenecks. You spend your time where you can most improve the system’s output. Sometimes you jump in and directly work on a specific task. But mostly you think about how to fix forward so the same types of challenges get handled better next time.

Every AI system’s bottleneck is a human input. Once you fix that, you can see the real bottlenecks, and that’s where it gets interesting.

What BOTTLENECK ANALYSIS looks like with a closed loop agent

What does a day in the life of a closed loop agent maintainer look like?

Everyone wants to build a Polymarket agent these days so lets use that as an example.

It picks a niche, makes predictions, places bets, and interacts with a Discord community to test hypotheses and gather intelligence.

You get it running and hit the first bottleneck, it’s not profitable. Is the sample size big enough? Are you picking events that are insider-driven and therefore unwinnable? Or is the calibration just off? Maybe you look at the calibration curve and find the agent is systematically overconfident on certain event types. You adjust baseline probabilities.

Profitability improves. The agent becomes more conservative, researches more deeply, and eventually finds high-conviction bets.

But now it’s spending too many research tokens. It’s investigating too many markets in parallel and running out of budget before completing analysis on enough of them. You run an 80/20 analysis on which research activities actually moved the needle on prediction accuracy and cut the rest.

Research efficiency improves. Now the agent is running out of markets in its niche. It’s putting up good predictions on a third of available markets, but needs to expand. So you broaden the niche.

All this work is iterative and high-leverage. You’re not doing the research yourself. You’re not picking the markets. You’re mainly tuning the machine. And you can still override individual predictions, pause certain bets, or add safeguards.

You’re defaulting to closed loop but maintaining quality.

So where can this be applied?

TWO kinds of closed loop agent problems

In practice, there are two categories where closed loop agents make sense that form a barbell.

The first is low-level automation where it’s not worth any human input. Bug triage and basic fixes for a codebase. Simple customer support for a product that can’t afford a support team. The agent does the best thing that can be automated, and you don’t look at it unless something breaks. Your only cost is tokens.

The second is the opposite end of the spectrum: core projects that are core your business. These are systems you’re trying to make closed loop because you want to improve the leverage and scale at which you can attack them.

Many problems won’t be closed loop and shouldn’t be. One-off tasks, work where you value the human voice, problems where human ideas are the whole point. This is not a hammer to apply to everything. But for the right problems at either end of the barbell, it becomes the natural way to work.

It also only gets better over time.

Closed loop agents are a HEDGE against the future

If model capabilities keep improving, and every indication is that they will, closed loop agents get easier and easier to build and more capable over time. The window between model versions is so compressed now that if you’re building exclusively for the capabilities of the current model, you’re probably building the wrong thing.

Closed loop agents are inherently forward-compatible. You’re not optimizing for a specific model’s quirks. You’re building a system where better models mean better performance on the objective, automatically.

This is why I’m starting to think about closed loop agents across several domains. The architecture forces you to ask the right question: can this problem be decomposed into an objective, a set of tools, and a feedback mechanism that doesn’t require me?

If the answer is yes, even partially, you should be building toward closed loop now.

I'm convinced closed loop agents will stop being just an attention grab.

For certain problems, they’ll just be how work gets done.

And it’s also a decidedly more fun way to use AI.

I constantly feel like I’m falling behind with AI but recently I wondered if that’s really true.

If we look closely at the people who appear furthest ahead, the ones spending the most time with the technology… you’ll find something surprising: they have very little idea what to actually do with it.

Five Ways People FAIL to use AI

Using AI to slack off

Picture someone with a 40-hour job that includes ~35 hours of real work and ~5 hours of coffee and Slack.

They’re now using ChatGPT and Claude and are left with 20 hours of real work and 20 hours of browsing Facebook.

This person has owned themselves in three ways simultaneously.

1/ They’ve demonstrated that a third of their job can be automated, which means there’s less of a need for the role.

2/ Other people are using the same tools to do better work, not just less work, so the automator’s unchanged output is actually falling behind relative to peers.

3/ They’re atrophying their brain. The premium on actually thinking, on engaged cognitive work, has gone up, not down. Every hour you outsource to AI without redirecting that time toward harder problems has a cumulative negative effect.

The more extreme version of this are the over-employment crowd, people secretly working four or five jobs in parallel. In a nihilistic world where everyone’s income is going to 0 in 2 years maybe they are nailing it but they are failing to raise their career floor in every other scenario.

At least these people are doing their jobs. The next group has turned AI exploration itself into the product.

LARPing

The AI influencer economy runs on a very specific loop: try the next model, post about it, declare it changes everything, repeat. Every new release, from 4.6 to 5.3 is the one that changes the game. They said it about every single model.

Between model drops, these creators fill time with flashy build projects. “I just built this in 5 hours.” “I just shipped this in 10 hours.” The problem is their entire audience consists of people who also want to build things in 10 hours. Nobody sticks around to use what was built.

What you’re watching is effectively an outsourced marketing arm of OpenAI and Anthropic. These creators are great at getting people to try the tools. But they monetize attention through courses, communities, subscriptions – all focused on telling other people to use AI. None of it is about doing anything with AI.

The mythical person who is quietly running away to build a $100M business is probably not posting about it.

The creators will tell you to “just start playing with the tools and you’ll figure it out.” But they’ve spent all this time playing with the tools, and they haven’t figured it out either.

Lazy solopreneurship

The third failure mode is the person who was always too lazy to build something and now uses AI to skip it entirely. X reply automators (with useless followers), Polymarket bot builders (unprofitable) and ebook generators (with 3 sales) are all in this category.

These people never did the deep work required to succeed at the thing they’re now automating.

If they had actually taken the time to write a book or post intelligently or specialize in some aspect of prediction markets, they might have been successful or at least learned some transferable skills. They hadn’t, so they reached for automation instead.

AI doesn’t fix the underlying deficit. Whatever revenue they generate temporarily will get commoditized fast, because the work they do now doesn’t make future work easier or better. There are no durable loops here.

Just vibe coding

The over-eager non-technical builder sees AI coding tools and thinks: finally, I can build an app. So they build a fitness tracker, a to-do list, a goal-setting app or a content generation tool.

Typically things they themselves would use, which makes them intellectually honest but constrained by a narrow view of what useful software looks like.

The fundamental mistake is believing that coding was the bottleneck. For most apps, the actual challenge is the growth loop. By removing the ability to produce the software, you haven’t removed the distribution challenge. You’ve just built a working product that nobody knows about.

“Gen AI” investments

Then there is the enterprise executive who read a McKinsey report on AI and decided to fully automate their call center. They signed a hard-to-reverse investment, hired a third-party contractor who promised to handle it, removed 500 people, and deployed AI agents.

Profit?

Nope.

All five groups share the SAME mistakes

What these five groups have in common is that they are all risk-averse and short-term oriented.

Which sounds counterintuitive, because some of these moves look bold. But automating your job, producing demo content, generating e-books, vibe-coding apps, and ripping out call centers are all attempts to capture value right now with minimal uncertainty. They’re low-risk, low-time-horizon plays.

Content creators will falsely tell you they have a view of the future: “you should be trying these tools out.” But they don’t have a thesis as to why. Their actual promise is that you’ll spend time with the tools and eventually figure out what to do. They haven’t figured it out either though, and they’ve spent way more time than you will.

This should be encouraging.

AI is not running away from you as fast as it seems. It’s a capability that is genuinely complicated to navigate, and it requires dimensions beyond technical proficiency to apply well.

Take RISKS, play LONG TERM GAMES

Spend a couple of weeks familiarizing yourself with the latest generation of tools. But don’t make it 50% of your week. You can rely on the content creators to surface what’s new and then intelligently switch over when a new great tool appears.

Once you have a sense of what’s possible, pick a narrow vertical and bet on it. Don’t start a YouTube channel of AI-generated Muppet sketches. Instead, take an almost science-fiction level North Star for where your vertical ends up in two years.

Assume models will be dramatically better. Then spend the next two years thinking about what’s possible in that world, building the components that are possible now.

You will spend more time in that future than anyone else. You’re learning things nobody else is learning. And that compounds into genuine domain advantage, the kind that can’t be replicated by someone who just picked up the tools last week.

There are better ideas out there

The lesson here is that you should use personal conviction but if you don’t have it, here are some business models I’m intrigued by.

Distribution-first companies

If the ability to respond to feature requests is now extremely fast, then the feature matrix is no longer an advantage. Being two years early to understanding that a customer needs a feature doesn’t matter if anyone can build it in a day. The real domain of competition becomes marketing and distribution. Companies that use the excitement around AI itself as a distribution mechanism, and maintain enough taste and speed to respond to customer needs, are well-positioned.

BridgeMind is the perfect example here: the content value of the journey is more valuable than the direct revenue from the apps.

Services

80% of the US economy is services. The things that make service companies great: business development, high-trust relationships, the “we’ll do anything for you” approach, fundamentally weren’t scalable. But now they are. And many service-oriented companies aren’t technical enough to exploit it.

So what does the service become? Do you automate the back-end and maintain the trust with the same number of people, doing 10x more per client? Or do you automate the whole thing and turn it into a product, competing with SaaS by having a broader, more fluid feature scope where clients can request modifications to the software itself?

Services aren’t glamorous, which is exactly why this category is exciting for the right teams.

Data-media hybrids

You can’t automate journalism or opinion writing. And pure data businesses, the ones that are essentially a bunch of SQL queries and scripts, are now easier to replicate than ever.

These two categories are the opposite of a media spectrum and are going to blend in ways they haven’t before.

One of the first verticals where AI can write genuinely useful content is where it can procure deterministic data that lives entirely online. A fully automated crypto data newsletter, for instance, is far more doable than most people realize.

You would compete with data companies by doing what they do more cheaply, incorporating data sources that are hard for humans to mine, and then leveraging the data to create media-driven distribution.

Honest exploration as a business model

A subtle counter-example to my thesis is Every. They are a good example of how to build an AI-native business while taking very little risk.

On the surface, they look like AI creators: exploring tools, telling you to use them. But they’re intellectually honest about not knowing what the tools are for yet, and they actually build and launch their experiments. They do enterprise coaching, instead of $100 courses.

Build stuff anyways

Notice that none of these categories are focused on what can be done cheaper with AI.

So what do you do if you don’t want to build a company?

One of the best things I ever did at a job was create a product. The less you think of your role as receiving inputs and delivering outputs, and the more you think of it as being an internal CEO of a new project, the more successful you’ll be.

Favor creative, value-driven projects where the upside is building something new and nobody else is looking (yet).

“You have one hour to sell me on AI”, I heard from a >$10M CEO last Friday morning.

What would you do?

..

..

..

I decided to go all in on Claude Cowork. Context is everything (he’s an ex-McKinsey consultant).

We spent:

• 10 minutes signing up to Claude and installing the app,

• 20 more minutes waiting for the Claude Cowork workspace to initialise,

• and then 20 minutes with Opus 4.6 on just two demos (I’ll tell you later what they were).

We blew through the Pro plan weekly quota almost immediately but what happened next likely changed the course of his company.

Or did it?

”Invest in AI!” proclaim the consultants, authors and influencers. Yet it’s not been obvious how.

Most advice boils down to “just use the tools” but there’s a difference between people who are bottlenecked by their token budget vs. your average Gen Z Tik Tok consumer that relies on ChatGPT for relationship advice.

It may work on a personal level but it’s even more confusing on an a company level.

What do you build?

What do you share?

How do you work around security risks?

What infrastructure do you depend on?

The smart money knows the answer boils down to “CLIs and skills”. CLIs allow you to leverage the immense power of coding agents to create repeated deterministic tools that are better than MCTs. We already explained that one.

And skills represent real institutional knowledge, intelligent special purpose agents you can wake up to do your bidding.

SKILLS

Skills are pre-packaged prompts used primarily by long-running agents like Claude Code and Codex. You can create them by putting a prompt in a folder with a SKILLS.md file or using a dedicated skill.

Skills can range from simple prompts that explain how to use a particular tool (e.g., a skill for creating PowerPoint presentations with the Claude pptx-tool) to playbooks to enlist Charlie Munger as your CFO:

Reusing prompts is the simplest use case for skills at a personal level.

Integrating tools is another. If you built a CLI (remember CLIs+skills?), a good way to integrate it in your workflow is to write a quick explainer for when and how to use the CLI.

Building playbooks and custom workflows is another. Everything from an editorial grammar check to a full smart contract auditing playbook can be expressed since skills plug into agents with planning and execution capabilities.

Skills REPLACE Custom Agents

The most powerful ways to think about skills is that they replace the need for custom agents in many cases (especially internally). I’ve built a number of custom agents directly and were bogged down by things like planning, having to be too prescriptive about workflow, providing access to basic tools (like providing Google Search API keys) and even chunking more complex tasks.

Ultimately, building custom agents directly isn’t worth it anymore unless you want to bundle them as part of an externally-facing product.

Skills, however, are the perfect way to create custom agents:

• You benefit immediately from all the system prompts, planning, context management, memory management, UI, access, tool and other capabilities of these remarkable general purpose agents (Code & Codex). Nothing needs to be rebuilt.

• Skills compose and can be used with one another in previously unanticipated use cases.

• Skills are readable by everyone in your organization (just like a playbook) whereas custom agents could have their prompt instructions hidden in Python/JavaScript files. Equally, they can be created by everyone in your organization as well.

But while skills can help accrue and share institutional AI knowledge, you also have to use them right.

It’s a SKILL ISSUE

Here are some conventions we are starting to apply to skills use at Auditless. Would love to hear if you see them differently.

Avoid sharing skills before you use them.

Skills are not a replacement for playbooks, standards and quality controls. You should have those in place already. Skills are a tool that helps you implement the standards. Skills can refer to these standards, however, so you may use them as a convenient place to align on issues but humans are responsible that the standards met, not agents.

It can also be tempting to write skills for everything that you may ever need to prompt for. But sharing a skill should represent some amount of iteration and confidence that the skill is effective at achieving a certain outcome. That can only come for repeated prompting and tuning.

Don’t rely on AI to generate your most important skills in their entirety.

There’s a temptation to let AI generate prompts and this is often applied to skills. Just remember that the AI-generated version of a skill is v0. The more you will use a skill, you’ll want to make sure that it represents the best possible playbook for implementing a certain task. It’s OK to jump in and edit a skill or even to do an eval from time to time in the same way you would do it for custom agents.

Use skills instead of AGENTS.md where possible.

Don’t put too much knowledge or procedural knowledge in your AGENTS.md (context) files. One of the primary benefits of skills is progressive disclosure and more efficient context utilisation.

Don’t build a skill for what could be a tool.

If you are doing something deterministic (file formatting, pulling a date, etc.), it is more efficient to use the right tool for it. If in doubt, ask your agent whether something could be built and if you’re non-technical, learn how to vibe code to build these tools. You’ll get increasingly comfortable doing so.

I mentioned two demos at the beginning. One was a research task on the CEO’s industry which pulled their top 20 clients and a summary of information that could only be extracted by carefully analyzing all their public statements and announcements. It was assembled in a neat Excel document. The latter was a 10-page presentation on the last 3 days of Australian politics.

Claude nailed both.

The Personal Computer Revolution allowed individuals to directly access software at home (and in more offices). Companies that understood how best to use software (or better yet develop it) thrived. A class of jobs we now call “knowledge work” emerged and over time became synonymous with effective computer use.

As an engineer, I needed to be proficient with an IDE. As a consultant, PowerPoint, Excel and the browser were more important.

When Ray Dalio wrote Principles, he described ways managers can think about the organization and their role in it.

He compared effective management to operating a machine.

Until now, most of us didn't have the privilege of a managerial position and by extension the ability to “build a machine”.

Now we do.

I believe “building the machine” and operating it is precisely the skill that matters as we shift to agentic work and both individuals and companies are now in a race to build the machines up.

Benedict Brady’s essay about Closing the Software Loop points to what the endgame looks like in software development: a system that continuously collects user feedback, interprets and implements it with dedicated agents.

There are FIVE STAGES to ASCEND

Agentic Coding is where people who tie their identity to being good developers often get stuck in. Humans review every piece of code and often focus excessively on form over function, getting very little done in the process.

Vibe Coding is the default non-technical entry point. Punch in a few prompts, get a UI back and keep prompting until you get loosely what you want. Who know’s if it’ll work?

Looping is where the prompter starts to carefully define not just the work itself but relevant acceptance criteria: tests, feedback loops, audit processes, etc. The work itself is often conducted in a repeated loop until all the acceptance criteria are fulfilled. For more details on “Ralphing” as one way to do it read our post.

Agentic Engineering is where the user starts to become an orchestrator of several different agents often working in parallel. There’s enough orchestration that the human can simply describe the problem to be solved (Job to be Done) and a planning agent, coding agent and review agent can work together to find a solution.

Finally, Closed Loop Engineering is the holy grail. This phase aims to entirely automate the job of 1 Product Manager and their entire Engineering team, a unit that when horizontally scaled can match the output of an entire software company. The hardest part here is finding ways to gather and sort through human feedback in an unexploitable way and translate it into roadmaps.

WANT to play a GAME?

I think the game is moving up these levels for as many projects that you are currently tackling and then do many that you can't even imagine.

Your goal is to maintain and enhance the closed-loop machine by observing its performance and ability to achieve goals you set for it. At times, you can insert yourself in the machine, but devoting too much time to that builds a machine that is necessarily bottlenecked by the human in the loop. A machine that doesn't work when you sleep.

I actually started sleeping better when I realized this. The existential dread started washing away a little bit. I started to see exactly how having some combination of software development, product management and consulting skills will help me build up this machine faster.

It would be a mistake to think that traditional jobs and skillsets are going to stay the same. It’s also a mistake to think that we will be vulnerable to immediate automation. The reality is in the middle: there is a new game, a new skill to master.

We are all building verticalized machines that aim to replace what each of us could have done with the help of an entire team.

Verticalized because your taste (not your skill) provides a quality ceiling on the machine’s output and your existing experience is an accelerator in getting the machine and its playbooks up to scratch.

Let’s play.

While most people on X using AI seem to be desperate for their first hit, Peter Steinberger is the opposite.

He sold a company called PSPPDFKit for 100M EUR, fell into an existential spiral and returned to build things with AI for the sheer joy of it.

His "CURRENT PROJECTS” list nearly exceeds my vertical screen real estate but of course it’s the little lobster on top that has been stealing the show: Clawdbot Openclaw.

X is now littered with articles that explain how to install the bot and give it access to everything. Did I fall for it? Kinda.

OK I DID DO IT

I went the Mac Mini route. But let me explain. It’s not because I'm a PM or an Apple boy (even though I'm both those things). I simply already had an M1 Mac Mini that I've been using as a sandbox.

I didn't install it for the obvious viral reasons either (I don't want an agent to surprise me with a daily app or handle my calendar).

In fact, these use cases don't really matter and most people are using Openclaw wrong. We can classify these as 3 big failure modes:

FAILURE MODE Nº1: PERSONAL ASSISTANT

I recently thought it would be pretty clever to ask Perplexity and Deep Research to look up “Japanese Toddler breakfasts” and curate something that was appropriate modulo my son’s allergies.

Needless to say it didn't work. In true 2010s fashion I ended up finding a Japanese food blogger and going through her breakfast category manually one-by-one.

If you've wondered why agents can now one shot entire applications but can't help me find recipes without some heavy prompt engineering, it makes sense. Math & coding have been prioritized in model training instead.

In AI Minimalism, I wrote down some ideas on how to use AI.

The automation ideas I'm seeing are just party tricks for your X followers and you don't need a Mac Mini bot for most of the high leverage improvements. There’s nothing wrong with using Openclaw as your personal assistant but try to prioritize high leverage automations.

FAILURE MODE Nº2: “HACK ME PLEASE”

If you don't see running your own Openclaw instance as a form of Tower Defense, you’re in trouble. Unfortunately the security guide is too long to read for most people so the least you can do is ask Openclaw to review it every time you try to increase the attack surface.

I've seen people do stupid things like install Openclaw straight on their main machine and give Openclaw access to their accounts and payment cards. Unfortunately these are only the obvious ways you can get in trouble. Prompt injection vulnerabilities and supply chain attacks are much harder to reason about and avoid.

So be safe out there.

FAILURE MODE Nº3: BURNING TOKENS

If used well Openclaw is a way to more efficiently use tokens, it can:

• Find ways to spend idle tokens at the end of each spend period;

• Build more disciplined routing paths for regular tasks;

• Use local models, etc.

But it’s even easier to do the opposite.

If you are already at capacity and don't effectively monetize your tokens, launching a cron job every 5 minutes will burn through your quota even faster. And instead of spending your tokens on useful coding tasks you will spend them on your morning news summary?

Careful model management is essential when using Openclaw because it has access to a lot of context and memories that one-off Codex/Claude jobs won't require.

I'm anticipating downgrading the primary driver model and instead making sure that Openclaw can route to smarter models depending on the task. When it comes to agentic workflows, the intern becomes the CEO?

So make sure you either have an infinite budget or teach Openclaw how to use your tokens well.

I think token budgeting will become an important capability for companies and individuals and agents can help.

WHY I'M USING OPENCLAW

For me, I would like:

• A sandbox for coding agents to run in a highly permissioned state;

• The ability to remotely remotely initiate work that may use coding CLIs, browsers or other things that are easier to install then find a tool for;

• The ability to schedule and trigger additional complex work.

These things are highly specific to my workflow and you may not care about them at all.

How do you figure out what matters for you?

While AI Minimalism is a good approach to consider one-off interventions with AI, an ongoing system like Openclaw is best used to tackle systemic bottlenecks.

Bottleneck analysis is the best framework to use with Openclaw. If you want it to mull things overnight, instead of making it generate ideas at random, make it think about moving that bottleneck instead.

And remember:

No business is ever truly capped if you’re open to releasing every bottleneck.

The first thing you should transform with AI is your life.

Not your Claude Code setup.

When everyone is talking about how to configure AI, do the opposite.

Simplify your life and create space for new ideas and workflows to emerge.

I’ve been spending the cold and dark winter months on something I call “AI minimalism” and want to inspire you to do the same.

WHAT IS AI MINIMALISM

James Clear defines minimalism as focusing on and committing to the fundamentals, instead of wasting time, money, or energy on details. The Minimalists define it as a lifestyle that helps people question what things add value to their lives. At its heart, minimalism is just a fancy word for prioritization and focus.

Minimalists generally have less possessions, less commitments, less digital distractions because they take time to curate and reduce things in their lives. It’s what they have more of that matters though: time, energy, money and more freedom.

Today, AI is not a force for simplifying our lives.

You wake up in the morning to your phone buzzing with notifications from an app someone vibe coded.

You go check up on your agents to see if your new app idea is finished. You glance over a couple completed research reports from Perplexity and Claude Cowork bookmarking 7 articles to read later. You open X and read the bot replies to your latest thread.

It’s 10am and you wonder if you should prioritize getting something done today but you get distracted by your WhatsApp group that talks about how to best use AI. The door bell rings and the Amazon driver shows up with some new clothes. Oh yeah, you asked Claude to recommend you some new hoodies. You throw them in a closet and go back to looking over some more completed jobs.

Everything in your house, calendar and desktop is built around the old way of doing things.

The best way to adapt your life around the possibilities of AI while thriving cognitively is a complete reset.

PRESS RESET

Resets aren’t easy when you already feel short on time, distracted, with a mountain of ideas and responsibilities on your shoulders. Less time to prioritize means less clarity on what you can even consider dropping.

AI minimalism uses AI as a counter-force against the very reality that technology and AI has imposed on us.

A reality of material and informational abundance at negative marginal returns on happiness.

AI minimalism uses AI to overcome psychological (procrastination), knowledge and resource gaps to simplifying your life.

Point your favorite AI tool as a weapon to free up 3 key resources: money, time and focus.

CANCEL YOUR SUBSCRIPTIONS

Struggling to afford the $200 Max plan for Claude?

All you need to do is export your transactions from all your bank accounts and send them to Claude for an audit.

Anything you don’t use? Cancel it.

Then there are subscriptions you use that you can simply rebuild and forget about.

Here’s a concrete example, we had a smart contract development guide up on the Auditless website hosted by Webflow CMS. I was able to get rid of that subscription completely by getting everything exported into a cross-linked GitHub markdown book. Took less than an hour. I’ve seen people go as far as building their own versions of Calendly.

Nowadays the most useful applications are already free.

People are switching from Notion (complex to access via API) to Obsidian (just files, easy for agents to work with).

Others are moving from Linear to GitHub issues (although to be fair Linear are doing their best to adapt and I think they’ll crack it).

CANCEL YOUR MEETINGS

Recurring meetings are subscriptions on your time. Each meeting also further divides your workday and splitting one focus block into two is horrible for productivity.

What meetings exist just to keep people informed? Create an automation that summarizes Slack conversations and sends a weekly PPP update to the other team. Or have everyone record a 10 minute voice note, transcribe and combine.

What meetings can be made shorter and more efficient by building dashboards that provide valuable context on an automated basis?

For the remaining meetings, work with other people to move them to the same day.

I have been using Tuesdays for recurring meetings, for example.

The simple change makes other days feel more focused and impactful.

SIMPLIFY EVERYTHING

Declutter your house by using AI to categorize items and figure out where they can be donated or sold.

Remove all notifications apart from the ones that are completely essential.

Leverage daily email summarization to stop yourself from checking it compulsively.

Automate an AI calendar audit to get ideas for how to free up your time further.

Delegate more tasks to team members by sharing your prompts with them (another thing I did and prompts > playbooks).

What about boring things you have to do regularly like grocery planning? Can you create a recurring task that does this for you?

Same for a workout plan.

Scan restaurant openings or find other ideas for going out.

Automate data gathering for your personal investments to facilitate regular reviews and reallocations.

Not all these tips need AI but you can use AI to get them done.

Literally copy & paste this essay into Claude and ask it to make sure you get all this done.

CAN I UPGRADE MY AI WORKFLOW NOW?

Do this first before you focus on your AI workflow.

A month from now, everyone will have moved on to the next tool for managing 5 different agents and you will have saved a couple cycles.

Instead you will have learned a thing or two about how to use AI to actually create measurable improvements to your life (time, money, focus).

And now you can reinvest that time to improve your setup.

Or even build out your terrible app idea, but please do that last.

Last year, I wrote about how composability has influenced the AI form factor, notably pushing more serious users towards CLIs:

Composability is what initially drew me to blockchains and I think is the reason why CLI interfaces have been so successful in coding.

There are an infinite amount of workflows which are enabled by the flexibility of these interfaces:

• Working remotely via SSH (e.g., working from an iPad);

• Running multiple agents in parallel;

• Configuring agents (simply using local Markdown files);

• Using the CLI as an MCP tool in another agent;

• Using the CLI for one-off tasks (even using it from other tools, bash scripts or other coding agents);

• etc.

As an unfortunate side effect, non-technical users were consequentially gate-kept from running agents autonomously.

All that changes with the release of Claude Cowork (available for Claude Max users in a research preview), which provides Claude Code-like agentic capabilities in an Electron Mac app.

It’s positioned as an experiment with the first iteration taking just 1.5 weeks to build according to Felix from the Anthropic technical team.

But the timing for the POC is perfect. Many people have already been diving into Claude Code over the holidays and will be relieved to find they won't have to learn Bash, Git, Markdown, tmux anymore to run long-running jobs.

Is this as good as Claude Code?

IT’S STILL JUST A UI

When I explained why CLIs are so powerful, it wasn't just that a CLI has access to the host’s computer. It's that the CLI can be invoked in various ways. “Ralphing” as described in last week’s issue is the perfect example of something that is easily done with Claude Code but impossible to do with Claude Cowork.

Moving from tool-enriched “chat”, a single-threaded conversation which is effectively just a context filling hack to “cowork”, an agentic loop is a big deal. Being able to run many cowork agents in parallel is also a really big deal. But one last important step is missing: programmatically invoking cowork agents.

Still, Claude Cowork can do a lot:

• It can access your computer’s file system;

• It can access your emails/calendar and use other common integrations AND it has the horsepower to actually process much more data in these applications to do things like comprehensive calendar audits or email summaries;

• It can even use a Chrome browser and interact with applications that may otherwise not have tools/apps. Even using tools to create presentations.

NON-CODE MEDIA

There is another reason why Claude Code has been so powerful and we should temper our expectations with Claude Cowork.

Coding is highly deterministic: you have tools to robustly check syntax, strict conventions for how to do things like writing comments and naming variables as well as ways to test for correctness.

Any vibe coder knows that you can iterate any app idea into submission by simply continuing to test & try the app.

The same is not true for getting AI to write coherent thoughts, generate beautiful images or completing long-standing research tasks. You can often get it to work but things don't get better with iteration the same way that code does. Often you get some sort of prompting insight that creates a step-change in output. And often that prompting insight is not necessarily applicable across domains.

Coding with agents is easy, nobody has yet figured out the playbook how to make world class presentations, spreadsheets, etc. I’m curious how that plays out. Couple of options:

1/ Operate on the prompt

The answer could be that you need Claude Cowork to iterate in a feedback loop on its own prompt as opposed to the artifact itself. So you'd iterate on the perfect spreadsheet writing prompt as opposed to tweaking the spreadsheet itself.

2/ Reduce everything to code

You can build presentations with Beamer, spreadsheets with Incremental and gifs with Processing. All in code. It’s possible that coding agents will become so powerful that it will be easier to let Claude Code write LaTeX for your presentation as opposed to letting it stumble around in Keynote or Google Slides.

It will be cool to see applications of AI extend beyond coding, slop content and reply bots on X into regular enterprise workflows but expect adoption to be a little slower. Developers love trying new tools way more than the average person in marketing & sales.

It’s hard to write a newsletter every week when you take two weeks off. My strategy for this has been to write shorter pieces focused on one simple idea.

This piece will talk about “Ralphing” (an effective way of using coding agents) and the impact it may have on software development.

Ralph is this idea by Geoffrey Huntley to run the same prompt repeatedly in a loop until it is done.

It's the ultimate way to double down on one-shotting coding tasks especially applicable for long-term tasks running overnight.

Typically when you try to one shot a task with Claude or another coding agent, it will give an earnest go over a number of iterations but may eventually stop. This could be because it wants to preserve credits, or it ran into a loop or it blew out its context or some other excuse.

“Ralph” circumvents that by repeating the Claude loop over and over with a fresh context every time (minus whatever code was created in the current folder). Doing this has been surprisingly effective. To speculate why:

• Performance naturally degrades at larger context windows so the hard reset is helpful;

• Using Ralph encourages user prompts to direct coding agents to write more self-contextualizing code (context is in the repo/comments) which is a good practice in general for performance and autonomous development; Note this implies your first Ralph may not work as well;

• Most people get in the way of agents instead of helping them effectively through a task.

Steve Yegge (yes! Steve) then followed this up with my favorite post of 2026 so far: Welcome to Gas Town. In this post, he predicts that software engineers will inevitably become orchestrators of several agents in parallel.

He also makes this image covering a concept I'd loosely described before in The Delegation Window and Limits to Vibe Coding.

Here is his framework:

First, you should locate yourself on the chart. What stage are you in your AI-assisted coding journey?

Stage 1: Zero or Near-Zero AI: maybe code completions, sometimes ask Chat questions

Stage 2: Coding agent in IDE, permissions turned on. A narrow coding agent in a sidebar asks your permission to run tools.

Stage 3: Agent in IDE, YOLO mode: Trust goes up. You turn off permissions, agent gets wider.

Stage 4: In IDE, wide agent: Your agent gradually grows to fill the screen. Code is just for diffs.

Stage 5: CLI, single agent. YOLO. Diffs scroll by. You may or may not look at them.

Stage 6: CLI, multi-agent, YOLO. You regularly use 3 to 5 parallel instances. You are very fast.

Stage 7: 10+ agents, hand-managed. You are starting to push the limits of hand-management.

Stage 8: Building your own orchestrator. You are on the frontier, automating your workflow.

The claim: it's inevitable that software engineers of the future have more in common with the Party Poker grinders of 2010 than anyone else.

As we all work our way up this curve, it does have a fundamental bottleneck that is not skill related: money.

At a certain point you need more cash to support workloads. So you need a job or business that generates ROI on the code you produce.

And this is an interesting phenomenon.

While coding agents are making coding much more accessible initially, in the future, software engineers may be stratified at very different tiers.

Some will be foot soldiers and others will be air force pilots, piloting $1M of credit a year.

Similar to how runaway companies that can get the LTV > CAC equation to work in their favor can hyper-scale, I suspect the coming decade will be dominated by companies and individuals who can get positive ROI on their coding credits.

Vitalik wrote something really interesting last week: Ethereum has finally become a high bandwidth network without sacrificing decentralization or consensus enabled by zkEVMs.

In parallel, apparently some of you still needed convincing? Opus 4.5 is the technical leap that finally sold the general public on coding agents.

I’ll spare you the subjective tweets (but I promise, it does one shot things pretty well) and just share this chart:

One of the things crypto does really well is turn nearly all aspects of finance/fintech into code writing (and auditing) problems:

• Need servers? Program crypto-economic incentives for other people to run nodes.

• Need guaranteed payment? Use an audited smart contract.

• Need liquidity? Write a Uniswap v4 hook and program it.

• Need to accept payments? Use a wallet SKD.

• Need to integrate another service? Just look at the ABI of the contract.

• etc.

And the happy accident of this decade is that coding agents are now really good becomes really convenient in this context.

Do you bet on the financial rails that turns everything into a coding problem or on the financial rails that still requires people, paperwork and process?

It’s probably not an accident that even the Ethereum ecosystem that has had its problems with core development velocity and resourcing is shipping faster than ever (especially in aggregate with the L2+ ecosystem).

Anything that is ultimately a computer science or coding problem is going to get solved faster and faster.

I've been making this shift on a personal tool level and wrote about how I stopped using no code tools even for agent building:

This was until recently my go-to driver. Getting the ease of a no-code tool but the ability to jump in, inspect each component and write custom ones is truly great. Things that are simple to write in code are not always easy to build in a No-code tool like n8n or Langflow.

But that’s not the biggest issue. The biggest issue is that you can’t get AI to help you build your pipelines.

I want to vibe code my AI pipeline.

I want to be building in Cursor or Windsurf, not dragging boxes around.

No-code pipelines are also not portable. These tools will evolve quickly and I expect to need to constantly migrate and upgrade our agents from one setup to another.

In the words of Douglas Engelbart: we need tools that help us build better tools.

While us developers may be early to the transition:

Companies won't sleep on it for too long.

In addition to project metrics, DeFiLlama also publishes public raise announcements. While the data is far from perfect, it presents a useful proxy on what areas both retail and institutional investors are allocating capital to.

I went through and labeled all the raises from the first half of December, here’s what that amounts to (amounts in $M).

I've taken the liberty of breaking these into subjective categories but there's plenty of overlap. What do you call a company that aggregates RWA perps? I've called them an aggregator by focusing perps/RWA categories for direct issuers but many alternative categorizations exist.

These raises include both VC rounds, token sales and pre-sales. They are also often dated on the day of the announcement, not on the day of the raise itself. VC fundraising in particular often gets tactically delayed.

PRIVACY + LEVERAGE + RWAS LEAD

The “hot” trends of the year are well represented at 45% of capital raised and followed by a trio of “evergreen” categories like aggregators, DEXes and CEXes.

Worth noting that these aren't blue ocean areas. Privacy is arguably the most technically complex area in crypto and one that hasn't had its breakout use case yet. You are also competing against companies like Aztec that have been in the market for a long time and have the best technical talent in crypto.

There is more to build in RWAs especially for non-crypto entrants but you need to slog through regulatory questions. RWA is more fintech than crypto.

And then there are perps (or other leverage tools, but mainly perps). On the surface this is the most accessible area for Founders to build in: you need protocol, good oracles and great UX. Worst case you may get challenged on an algorithm that you don't even need to open source. On the flip side you have to compete with Hyperliquid and unless you are doing something net new like Pendle Boros, it's going to be difficult.

AI AND DATA

Only 5% of capital going to AI companies was surprising and the most exciting companies in this cohort are data companies like Surf, a version of Perplexity that isn't bad at crypto searches which recently raised $15M.

The decentralized model space retains its challenges (hard to beat commercial models and right now that’s what the customer wants). The agentic economy is another rising narrative but it’s led by a set of interoperable standards rather than breakout companies, so we’ll have to see what it amounts to. Thank God agentic launchpads have cooled off and overall it’s nice to see crypto exerting more caution about AI than the wider market (I believe rightfully so).

THE REST

Out of the other categories, Payments is one I'd expect to keep growing and see real adoption. I'm not too optimistic by the competitive CEX/DEX/Wallet products I saw as existing players have strong moats and actually raising their fees (Coinbase, Uniswap, etc.) which only happens from a position of strength.

The saddest thing perhaps is that I didn't have to include an “Other” or “Rogue” category to cover new types of products or innovations. SpaceComputer is a cool idea to build a compute cluster in space but it unfortunately nearly escaped the December list with the funding announcement in November 26.

There’s a lot of opportunity for new Founders to push the industry forward but I'm not excited by these partiular areas. There’s so much opportunity in crypto services, crypto-distributed network states and futarchy infrastructure and doesn't seem like investors have enough meaningful options to support these areas right now.

P.S. I tried this new format since there wasn’t an obvious project I wanted to cover. Did you find this analysis interesting? Is it something we should revisit in the future? Or would you rather see me pick and cover one of these projects?

My belief in crypto could be directly attributed to a project I did at McKinsey. I had to work with a retail bank and no that doesn’t mean in the HQ (although there were plenty of those projects as well). I got to physically work in a branch for 2 months.

As part of that I got to sit in and listen on advisory conversations. I saw sad situations of customers requesting to consolidate their debt, a bank adviser telling them to not do so because interest rate would be higher than any of the individual loans and still having to complete the request because the customer thought it was “simpler” to have one big loan.

I'm a fan of Coinbase and hold the stock but I worry about the direction they are headed in post System Update (their latest release event). I don't think Coinbase appreciates how much trust people will place in their advice. Nevertheless, there are some products here to get excited about so lets go through it all.

BRIAN IS QUESTIONING CONTENT COINS

We've written about the complex reality of Content Coins already. It's not entirely right to think of them as being antithetical to non-custodial trading. As we show in the piece, Content Coins are in fact the most trading-oriented implementation of social media possible onchain.

But Brian highlights a UX problem: Base App replaced Coinbase Wallet instead of evolving it.

The app now allows you to switch back-and-forth between the two apps but if you want to create private claim wallets in Base App, it’ll be annoying having a social profile and content coin forcibly attached to each of your accounts.

COINBASE ADVISOR RECOMMENDS TRADES

There’s no better way to summarize the current regulatory climate than Coinbase releasing an app which gives AI generated investment recommendations and lets you execute trades directly.

Here’s a portfolio that “managers with combined 75 years of experience” would recommend. A combination of:

• Coinbase held BTC which offers 0% yield and none of the noncustodial benefits of holding BTC;

• Coinbase held ETH which offers an abysmal 1.7% staking yield and a 1% withdrawal penalty;

• Coinbase held USDC which offers 0% yield unless you pay for a Coinbase One subscription (only One subscribers will get access to the Advisor anyways).

This is obviously just a repackaging of Coinbase’s product line instead of a serious portfolio adviser so be warned.

You could argue that at least Coinbase is a pretty good custodian for your coins and the price is giving them all the yield.

It gets worse though: Coinbase Advisor will actively recommend you to take positions in prediction markets. I still believe people under-estimate how much of these prices are driven by insiders.

During the System Update call, Coinbase seemed to suggest that there were many good options to speculate on AI development through prediction markets or semi-conductor companies at a time when prediction markets are rampant with insider trading and many consider AI equities to be overpriced.

I think a company that prides itself on making the financial world more transparent and accessible should also seek to innovate on the way financial products are marketed. Vanguard is still top of class here.

THE FIRST M0 COMPETITOR

We covered M0 recently, showcasing its unique protocol design and early response to the demand of custom issued stablecoins:

Aggregators, games and other applications which would naturally end up facilitating user wallets and payments now have an easy pathway to create their own stablecoins without having to worry about the issuance side at all.

The most elegant thing about M0 is that they’ve abstracted the complexity of issuance and cross-stablecoin liquidity behind a single ERC20 token and a set of smart contracts you can deploy and use.

Rollup stacks created a proliferation of L2s launched by apps which allowed the apps to access sequencer fees as a revenue stream. Similarly, M0 will allow more apps to issue their own stablecoins to reclaim yield.

Based on the post, it seems like the tradeoff is that you get additional distribution from Coinbase (e.g., tradability in their venues) and can manage the supply in a simple, custodial way with Coinbase Prime. Whereas M0 allows more onchain native distribution and more programmability in how you configure the stablecoin, allowing rewards to be sent onchain, distributed to users, etc. Based on that I think Coinbase/M0 will be competing in two distinct (yet somewhat overlapping) niches.

TOKENIZED STOCKS ARE HERE

Stock tokens have clear potential of dwarfing every other crypto asset in scale, have clear benefits (24/7 markets, lower trading costs) and a launch by Coinbase suggests that we are ready. This is a big moment and the only thing to look out for is volume in onchain venues like Uniswap vs. off chain.

All in all, Coinbase just made another step forward to becoming a bank: a place that you deeply trust and never change but monetizes you pretty aggressively through subscriptions, high fees and low yields.

On a positive note, they are making meaningful technical strides along the way:

• They could have just added stocks and started competing with Robinhood but instead figured out how to tokenize them;

• They could have ignored Base App in favor of pushing Coinbase as the super app but instead they made Base App available to everyone and are trying to simultaneously build the best custodial and non-custodial experience out there;

• They built an AI advisor that (while not perfect) recommends exposure to crypto-native assets while showcasing how crypto is connects to the real world (through prediction markets and news).

I just hope they lean on helping users save more money as opposed to trading it all away.

“We are not just here to build explorers”

This is the line I heard on the Trustless Agents Ethereum Community Call and it encapsulated my confusion with the new ERC-8004 standard.

This is a new standard for onchain agent registries that has gained popularity following the Trustless Agent Day at DevConnect. It has been compared with with Anthropic’s MCT, Google’s A2A and Coinbase's x402.

Here’s a summary from Hubble:

New chains and infrastructure teams in need of a growth narrative are latching on to ERC-8004 and everyone in crypto is writing explainers the “trustless agentic economy”.

I don't get it.

The messaging around ERC-8004 is really vague:

• “It enables onchain economies”;

• “It enables trustless agents”;

• “It enables agentic finance”.

This doesn't just come from ChatGPT generated blog posts, this language is used in the official roadmap (NOTE: finding actual use cases is a task for December in small print):

We already learned the lesson with stablecoins. New entrants like Stripe and Robinhood aren't talking about “enabling DeFi”, they are talking about how stablecoins are cheaper/faster to transact with, easier to access, available 24/7, etc. I wrote about this here:

The other pissed off group are people who thought they would win crypto just because they were early. They expected the complete dissolution of tradfi and hoped that their little startup will end up hiring JP Morgan VPs as they flee to become degens. Instead they now see they will be crushed by tradfi companies like Robinhood who have brand recognition, regulatory moats, people that actually work hard, a focus on user experience, trust with their customer and so on. It’s over.

I think crypto agents may be suffering the same fate. People think that just identifying and riding the trend is enough to win; in reality that’s far from the case.

Especially in the early days of a new trend forming, you need to cut through the crap – that’s what I'm here to do now.

WHAT IS ERC-8004

ERC-8004 establishes an onchain registry for agents.

It’s a smart contract registry of agents with three components.

Identity Registry

Basically points to a registration document for each agent, which looks like this:

One cool thing about the registry is that it is an ERC721 contract (read: NFT collection) where each agent corresponds to an individual NFT.

In addition, arbitrary metadata can be set on an agent such as additional wallet addresses.

Reputation Registry

The reputation registry allows consumers of the agent to submit “feedback” and have that automatically be aggregated.

The feedback can only be submitted by a real client, somebody who asked the agent to complete a task. The feedback includes a 0-100 score, two tags and a link to a JSON file for additional information.

Any client can give feedback by calling giveFeedback or revoke it by calling revokeFeedback.

The standard motivates building custom reputation systems that process and aggregate feedback into reputation scores.

This is an example of what a feedback file could look like:

Validation Registry

These are hooks to use deterministic validation methods to validate task outputs.

The agent owner calls validationRequest to request a specific validator to validate a specific set of tasks and the validator responds by calling validationResponse, providing a score from 0-100.

This is comparable to task validation in Bittensor subnets (although those are not specific to agents).

THE MANY FLAWS OF ONCHAIN REGISTRIES

This all sounds good in theory.

If you think of this as an onchain marketplace for agents like Uber, we have some of the key ingredients:

• The profile of an agent (the rider);

• Their “rating” in the form of reputation;

• A way to validate their outputs (equivalent to GPS-based rider tracking in Uber.

But there are two huge challenges: getting people to use the open marketplace and security.

In the early days of crypto, there was an idea called token curated registries which would use a token to incentivize people to maintain registries (e.g., a top 100 movie list). People even wrote papers supposedly proving that there would be enough of an incentive for theses lists to get accurately maintained.

Here’s a line that people tended to use:

The intrinsic tokens of token-curated registries are necessary elements of self-sustaining systems which are public utilities. Token-curated registries are peak predators of capitalism that perform a useful function at the lowest possible marginal cost.

In practice, none of the token-curated registries got any adoption.

Agent marketplaces will absolutely exist but why would they publish all their data in the open?

In almost any market, the winning marketplace is one that offers significant additional value or has the strongest network effect. Publishing your “supply” into an open registry removes the network effect moat altogether. The Apple App Store proposes native access to the most affluent large mobile phone user base AND comes with a custom programming language and SDK specialized for phone development.

Do you really believe the winning agentic marketplace will be onchain vs. from OpenAI?

One argument for using these registries is that there are people who care to consume agents “trustlessly” and don't want to go through a centralized platform.

But this registry:

• Doesn't host the agents;

• Doesn't hold the agent developers legally accountable in any way;

• Doesn't provide a way to validate if the feedback provided is actually valid (and not provided by competing agents, for example).

The best marketplaces are not trustless, they provide a lot of tools to help you trust the other marketplace participants.

Which brings us to the security issue. None of the underlying mechanisms is bulletproof. Restricting feedback to only be submitted by verified clients is a very weak and clearly insufficient form of validation.

Lots of additional infrastructure needs to be built on top here to actually produce a robust reputation oracle for agents.

And even then, the agent developer can change the source code of their agent at any point (unless you are using specific validators via TEEs).

THE BULL SIDE

I guess my point isn't that these registries are completely useless, it’s just that we are too early in the agent lifecycle for people to pick agents from an onchain marketplace.

Agents hiring agents

There are use cases that could benefit from an onchain registry such as agents discovering other agents and then enlisting them to complete tasks.

Tokenized marketplaces

Perhaps some protocols figure out a way to make open marketplaces work through tokenomics (such as rewarding agents for receiving positive reputation).

Not just agents

Finally, I compared the interface before to the Bittensor validation process. I think this standard is broad enough to be adopted in use cases that aren't just agents (or equivalently, the interface for an agent is broad enough to interface over programs, data and other non-agentic entities).

I think this is a good long-term discussion to have but right now we just need to solve real problems like delivering yield to users through agents and making the first truly useful agents.

2025 is coming to a quiet and painful end and we’re questioning everything again starting from blue chip chains like Optimism and Bitcoin to AI.

Here are some predictions on areas that I care about (be warned: no correlation to the areas that actually matter).

Looking back on this should be a good exercise next year and in the worst case a reminder to you that you shouldn't take writers on the internet too seriously.

1. Fintech L1s accrue no value to $ETH. Ethereum becomes a marketing term and the “Ethereum community” starts caring about the asset again.

2. More stablecoins than L2s. Same mechanism but way easier to launch. In general, the industry learns to monetize TVL as much as it monetizes transaction volume.

3. Rollup cluster meta is over. The idea of giving away significant sequencer fees for rollup-to-rollup connectivity goes away/is commoditized. We see major exits from rollup clusters.

4. Privacy retracts. I think we are still a couple iterations away from mass adopted applications of privacy tech so while it’s great to see the the value spike I think it is a little premature.

5. Onchain equivalents to ETFs and end of retail farming. Via vehicles like gtUSDa users shift from farming specific vaults to depositing into managed vaults for each collateral asset they care about (USDC, ETH, etc.).

6. A “trustless asset management” protocol gets exploited for >$100M. Protocol architecture starts to matter again. Not one that I’m rooting for but seems highly likely given the growth of the segment/participants pivoting into earn programs and the shortcuts protocols are taking. We’ve spent blood, sweat and tears building the safest possible architecture for onchain asset management and remain amazed at the volume deposited in vaults that are effectively multi-sigs.

7. First Bitcoin L2s launch. Nobody’s watching and there’s still a lot of narrative confusion around this but there are several teams behind the scenes working on Bitcoin L2s. The launches begin to challenge the story that BTC is “just a memecoin” or lindy-coin.

8. The first non-USD stablecoins see adoption. I don't think anybody particularly cares about GBP stable coins but CHF or SGD would be another matter entirely. More to explore there.

9. Rollups compete on apps. “General rollups” without a killer app struggle to compete on transaction volume.

10. AI replaces communities and social media for asset discovery. Memecoin communities have peaked.

11. Traditional NFT collections do not bounce back next year. However, expect to see more tokenization via ERC721 even if it’s just Pokemon cards.

12. A major celebrity launches a Creator Coin on Base, then everyone forgets about Creator Coins. I think Creator Coins are going to follow a similar trajectory to memecoins and friend.tech. The negative flywheel effect when early holders start to actually lose money does not facilitate long-term patronage in current iterations.

13. Major issuers develop their own custom Uniswap v4 pools and own their liquidity. Focused on RWA/KYC use cases initially.

14. Exchanges and other integrators rush to recommend onchain products. Now that they've figured out ways to monetize them (own chains, own stablecoins, deals, etc.).

15. Futarchy overtakes Retro Funding. Both through launchpads and campaign platforms like Butter.

16. Launching tokens is cool again. With regulatory support and better launchpads, more teams are emboldened to launch their tokens. This is somewhat dependent on asset prices in 2026 but we should also see teams launching tokens as a source of early funding through futarchy platforms.

17. “delegates” become a dying breed. The power of DAOs and their delegates shrinks substantially now that they are not a legal requirement. Protocols that didn't reversibly commit to a DAO model now have a competitive advantage in growth and momentum.

18. Network states issue real assets and become one of the most talked about things in crypto (for a week). Whether via real estate ownership or token issuance, it’s time to take the next step.

19. AI vulnerability detection starts working. First tools that can find independent vulnerabilities in smart contracts are established. However, they still don't replace audits and teams that rely entirely on AI auditing get burned.

20. 1-2 major prediction market betting categories disappear. We discover that insider incentives lead to unfair markets and platforms self-regulate or get regulated.

Uniswap just launched the auction to fix all auctions which they call Continuous Clearing Auctions.

They make some bold claims in the Whitepaper including that this mechanism could replace how IPO launches happen in the future.

Even cooler: the mechanism is immediately being put to use by the Aztec team which has decided to launch the $AZTEC token during a privacy boom (privacy-related tokens like Zcash are experiencing an unprecedented bull run).

HOW IT WORKS

The key idea of the auction system is that each bid is spread out over the whole time horizon of the auction. Here’s how things work step-by-step.

1/ The token issuer creates the auction and decides how many tokens will be available in each block.

This forms a release schedule (think of it as a supply curve). You can add a number of extra parameters too like custom code for a validation hook.

2/ Each block has a single clearing price that matches aggregate demand and supply.

Users express bids in the form of simple intents:

Once a block clears at a certain price, the next block can only clear at the same price or higher.

3/ Bids are spread out over multiple blocks.

Once you make a bid, the amount is spread out over the range of future blocks, meaning that you are automatically bidding on every follow-up block as well.

4/ When a target amount is cleared, the auction is graduated. If it doesn't eventually graduate, all bids are returned.

The main benefit this provides is that sniping is not possible.

• Late snipping is not possible. Delaying bidding only makes sense if you think you will get more information. You’ll almost certainly be buying at higher prices later on.

• Early sniping is not possible either since you can't concentrate your demand in early blocks.

CCA COMPARES FAVORABLY BUT HAS DRAWBACKS

CCA favorably to other auction models, here’s a not so scientific overview:

The main benefits of CCA are that it is resistant to timing attacks, encourages demand to form early and allows for the price to develop smoothly.

It’s not perfect, however. Here are some possible issues.

Under-bidding

However, like unsealed one-shot auctions, sophisticated market participants will be able to under bid (residual bid shading).

Secondary markets

Another issue I don't see talked about is that people could build secondary liquidity markets to complement the auction mechanism and allow participants to sell or hedge their future bids or won holdings (especially post-graduation).

This also could somehow invalidate the validation hooks, allowing users to create proxy accounts that are validated to indirectly support bids from non-validated users.

Attention minimizing

Finally, something Tarun suggested is that attention is an important byproduct of liquidity bootstrapping and in some ways while CCA aims to create as little trading activity as possible, that can be undesirable for discovery especially for “Veblen” goods. From the Doppler whitepaper:

Pre-pool sniping

This is something I'd like to see onchain evidence for but the idea that sniping is not possible during the auction doesn't mean that the transition to a liquidity pool will be smooth.

Some form of sniping may still be possible due to the way that the auction will convert into a Uniswap v4 pool with price equal to the last clearing price.

You could imagine a strategy where automated bidders would make sure to get in early in the auction and then be first to sell at a later clearing price. If they can guarantee early swapping they take very little risk that price actually went up during the auction.

Currently liquidity bootstrappers are encouraged to work with the Uniswap team to plan the launch carefully and they should definitely do so to manage this risk.

HOW TO PLAY THE $AZTEC SALE

As for the game theory of buying $AZTEC, it’s not obvious?

If you’re pricing the token on fundamentals and want to be a long-term holder, on the surface the CCA model encourages you to get into the auction as early as possible and with size.

If you are a larger participant (whale), you may want to under-bid to avoid pushing the market upwards too much if you think this will buy you more depth in the long run.

Since the CCA auction locks in a price floor at the very beginning, you should also consider the real possibility that the price will drop post-auction.

If you’re planning on pricing the token based on narrative evolution, this would be another reason to delay or split your bids early on to understand demand better.

Finally, if you want to snipe the auction, you need to study the specific way the initial Uniswap v4 pool will be seeded and think at what price you'd reasonably be able to sell tokens back into the pool once the auction concludes.

In conclusion, I think CCA does not entirely remove skill and technology as a factor in the auction process. Clearly, even if you desire to be a long-term holder, there are quite a few ways you can end up paying more for the same amount of tokens.

However, with the right post-auction pool design, I think it could require attackers to take on more price risk than they'd like and I expect this alone would significantly mitigate the impact of sniping.

More broadly, CCA opens up an exciting new design space in block-by-block auctions and the protocol design applications beyond token sales are promising.

Uniswap Labs is one of the few teams left in DeFi that can and do build net new protocols and I'm looking forward to the next one.

Crypto Twitter didn't react well to the $JESSE launch:

(And the above tweet was among the more reasonable and grounded complaints that I found.)

Others pointed out:

• Bad timing (the launch came too soon after David Phelps post complaining that Base is overly focused on Creator Coins);

• Extraction concerns, basically claiming that $JESSE extracted significant transaction fees from the sale;

• Sniping. Since $JESSE used the Zora x Doppler bonding curve auction, it inadvertently attracted snipers.

But I don't buy into these concerns.

Timing was a little unfortunate but I'm guessing Jesse had already been planning to launch for a while and before his birthday.

It wasn't extractive either, during his Birthday stream he was able to reinvest fees into other creators on Base. He claimed that he doesn't intend to ever sell the tokens.

Finally, there was a pretty good conversation with Doppler and 11AM that addresses sniping:

We’ll go more into depth on the pros/cons of different auction mechanisms in next week’s issue but Austin has thought about auctions more than every single person who complained about sniping on X.

So if not for malicious reasons, why would Jesse do this?

THE REAL REASON CREATOR COINS ARE PUSHED

Rollups make most of their sequencer fees from trading.

To date, Base has generated more revenue from memecoin trading than any other activity.

Issuance of new tokens and resulting speculative volume is a major fee driver.

It's likely that Base is spending more than ever on the core team, grants, events, own apps like Base App and supporting Founders but none of that is materially growing Base’s contribution to Coinbase’s balance sheet as a rollup.

Creator Coins and Content Coins are a very clever solution to this problem:

• They create even more tokens than memecoins (issuance is a primary battleground for rollups);

• They generate trading/speculation activity;

• They are built to translate attention into onchain fees. Anything that is going viral by definition will have a Content Coin attached to it;

• Even more so than memecoins, they don't require any underlying economic activity, nor a community, or promises.

While gas spikes are framed as a negative by users, creating excess demand for block space is actually success.

No other form of creator monetization would have the same effect:

• Payments (e.g., donations) of any kind don't have enough volume let alone payments to creators.

• Rewards. Base do use these to seed Base App but rewards have negligible impact on revenue.

• Advertising. Onchain advertising doesn't really exist to speak about sequencer fees.

• etc.

ARE THEY ACTUALLY GOOD?

It’s clear why Creator Coins make sense as a focus area for Base, are they actually the best mechanism for users and creators?

The flywheel is simple:

• If you post, it’s a token. You own 1% of the supply.

• Each Content Coin can only be purchased via your Creator Coin. If you launch a Creator Coin, you own 50% of the supply (vested).

• Demand for your Content Coins naturally creates demand for your Creator Coins. This aligns incentives so that you are motivated to create great content and benefit from ownership of your Creator Coin and the trading fees.

At some level, Content Coins behave like a Patreon membership. If you buy $1000 worth of a Content Coin, your opportunity cost is investing that $1000 to get yield at market rates elsewhere. This forgone yield operates as a creator subscription. In exchange, the creator may decide to reward you for holding the tokens. These rewards could be tiered like Patreon, scaled or random (lottery).

However, the creator doesn't directly receive the subscription unless they dispose of their Creator Coins. So even though your costs are in the form of a yield subscription, not all of that is effectively transferred to the creator you are supporting unless they “rug you”, Jesse pointed out this issue as well:

In addition, Content Coins behave like a fan collectible. The more popular an artist becomes, the more valuable the rewards they can offer to “subscribers”. Therefore, Content Coins have a speculative element. Even if you don't care about supporting the artist or the rewards, you could purchase a Content Coin just to speculate on the value of the rewards (material or immaterial) associated with holding the Content Coin later. This can be equivalent to buying a first edition CD from an artist you enjoy, you may be able to resell it later for more value.

The (significant) downside of this aspect of Creator Coins is that they are well and truly a financial instrument and their market may have institutional participants that have more sophisticated tools than the average fan.

It takes a real fan to recognize, preserve and invest in vintage CDs or swag. It just takes a smart trader to snipe or otherwise speculate on Content Coins.

A one-off slippage donation upon selling. Finally, to “exit” your membership you have to sell the tokens and will incur slippage. The less liquid the Creator Coin AND ironically the more you were contributing, the more slippage you will incur. Content Coins penalize the biggest patrons in this way.

THE HOUSE WINS

My ultimate problem with the Creator Coin model is that it attempts to combine patronage and curation in order to maximize trading volume and gets the worst of both worlds?

• True patrons have to deal with price volatility, adversarial market participants, trading taxes, etc.

• Curators have little guarantees around future rewards because the Creator Coin is not tied to creator equity or other value streams explicitly. Curators are speculating on unknown demand for future unknown rewards.

• It front-loads revenue generation for the creator, likely resulting in a larger amount of trading fees in the initial price discovery phase but unclear whether volume is sustainable (we will see with $JESSE). This doesn't result in great incentive alignment between the creator and the holders.

The result is that the underlying blockchain and trading venue (in this case Uniswap) extracts way more than a payment solution would in a simple membership model.

You could argue that the curation markets fulfill an additional function but the two cannot be cleanly separated.

As a polar opposite, we can study Craig Mod, he built his own membership stack, focused on keeping the setup as lean as possible and it works.

He can also proudly say that nobody lost their hard earned money trying to support him.

What I'm drawn to in Craig’s model is that the focus is on what is being created (books) as opposed to the content or the creator itself.

Personally, I think a creator economy that is focused on engagement is inferior to one that is focused around a real exchange of value and just using the content as a discovery mechanism and a way to work-in-public. A free tier of sorts.

I also think that these issues could be fixed and no doubt that the Zora and Base teams are working hard at it.

At least Creator Coins represent a real attempt to try something new in creator monetization and even if it doesn't prove to be the best model, it’s worth a go.