You Should Use The OpenAI SDK
Strategy, utility & implications.
The “smart money” in crypto is already using AI extensively.
Small applications like Georgios’ podcast synthesizer make a big difference:
But apart from AI agent issuance and trailblazers like Kaito, not enough companies are using AI for workflow elimination and scaling.
So this post will serve 3 goals:
Predict how enterprise value will be created over the next 3-5 years and explain why this hasn't been obvious yet;
Show how the new OpenAI SDK can drive that;
Explain its strategic context and OpenAI’s new plan.
AGENTS = ENTERPRISE VALUE
Following the DeepSeek launch, many speculated that OpenAI will need to shift their strategy or risk getting commoditized.
Since then:
OpenAI launched Deep Research, an agent that can complete multi-step internet-based research tasks;
Last week OpenAI shattered the AI equivalent of “fat protocol thesis” by releasing their new SDK;
The first two clearly center around the creation of agents.
Agents are pure, liquid enterprise value yet we still underestimate them.
We underestimate the power of agents for three broad reasons:
We undervalue the impact of chain-of-thought. Most people’s lived experience with AI chat interfaces doesn't blow them away. Explained in Why You Suck at ChatGPT. Effectively single-shot prompting is a very atomic and raw way to use LLMs that has significantly lower IQ than agentic workflows;
We don't appreciate volume multiplication. Our precious workflows are like luxury garments during the textile revolution. Luxury manufacturers didn't see the value of automation because the product quality was lower. Why automate something I can already produce better? What they didn't appreciate was the additional gains in customization, distribution and sheer volume. Imagine you work in BD and send 40 emails a week. ChatGPT can't write those emails better. But it could do 1000 emails and follow-ups on your behalf OR pull in research data on all your prospects OR create a custom 20-page report of your product that is tailored to each recipient.
We under appreciate the focus benefit of full task elimination. Despite all the tooling we have access to, I don't believe we are an order of magnitude more productive than people were crafting spreadsheets 10 years ago. Our work lives are littered with micro-tasks and social media replies that destroy our ability to focus. Being able to fully outsource a task to an agent will bring unexpected impacts in other areas of our jobs.
The 2nd effect in particular (volume multiplication) is why agentic workflows will be the source of significant enterprise value growth over the next 5 years.
Agents will:
Replace most productized services and eventually swallow the B2B service industry;
Raid job boards and entirely replace jobs at higher and higher salaries. At lower levels, every job applications will be replaced by 100 working agents very quickly at less spend;
Free up executives to focus on higher impact problems and solve them faster;
etc.
ENTER THE SDK
That brings us to the SDK.
I'm particularly excited because I've had some personal experience with the alternatives while I've been tinkering with simple AI workflows and wondering where to set them up.
There hasn't been a good alternative, until now.
LangChain?
I'm normally a buy vs. build person but LangChain is really difficult to use.
It is a cludge of incompatible abstractions and tools with limited utility like wrappers over search that AI could have written for you.
It is so bad that it's the first thing in a long time that reminds me of the R package ecosystem.
No-code or Low-code tools?
My favorite is n8n. No-code tools are awesome for people who don't code. Great for debugging and building pipelines quickly but they don't scale as you have to create boxes and arrows with your mouse!
As for low-code, my favorite tool has been LangFlow.
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.
The OpenAI SDK
The SDK is perfect:
The abstraction set is simple.
It will be compatible with other APIs following the same interface.
It has debugging and tracing built in! I was literally contemplating writing something like this and now I don't have to.
the OpenAI SDK bundles tools, even advanced ones like OpenAI’s browser controller. The problem when jumping out of ChatGPT or out of a No-Code tool is that you have to write your own tools. This is a pain: I've used the Google Search API (one of the simplest tools) directly in LangFlow and it has all sorts of weird quirks like limitations on number of results.
I know this may not be obvious yet but The OpenAI SDK (or something like it) is the best thing to use when building agents today.
Please do try the alternatives for a few days and let me know in a comment if you disagree.
THE STRATEGY
Is OpenAI trying to vertically integrate and make it harder to switch to other models?
Well, yes, but if the prospect of vertical integration makes them build better models, I'm happy to pay more for the underlying LLM calls!
They are also setting the standard for what open source model-agnostic frameworks should look like and hopefully paving the way for us to move away from LangChain.
(Yes, you could build your own Linux based fully open source framework with your own GPUs and your own local DeepSeek and just pay utility bills. But that costs you more in time than what you get.)
Switching costs don't really exist in a world where you can ask your IDE to fully rewrite your use of any one SDK to another.
And yes, the SDK doesn't do a 1000 things that LangChain does out of the box.
But AI can write those snippets anyways, in the style and context of your project. You only need a small set of primitives to work with.
No-code may even win eventually. I expect we will be building workflows by voice soon enough.
But right now No-code tools actually make you do more work than the coding tools do.
With the SDK launch, OpenAI’s biggest liability (investing significant compute in models that other people use to train theirs) becomes a tool to “commoditize the complement”.
Complement to what exactly?
Their new, much bigger business: monetizing the application layer (agents).
I'm sold and you should be too.