Anthropic Just Told You to Stop Using Agent Frameworks

Anthropic published a piece today on building effective agents, and buried in the practical advice section is a sentence that should make a few VC-backed developer tool companies nervous:

We suggest that developers start by using LLM APIs directly: many patterns can be implemented in a few lines of code. If you do use a framework, ensure you understand the underlying code. Incorrect assumptions about what's under the hood are a common source of customer error.

The people who built the model are telling you to talk to the model. Not through LangChain. Not through a multi-agent orchestration platform with a drag-and-drop interface and a $20/month pro tier. Directly.

This is the kind of advice that sounds obvious until you look at the ecosystem it's being published into — which currently contains roughly four hundred tools designed to abstract away the part that Anthropic is saying you should not abstract away.

The rest of the piece is good and worth reading slowly. The core taxonomy is: workflows (predetermined code paths, LLM fills in steps) versus agents (LLM directs its own process dynamically). Most things that get called agents are workflows. Most things that get called workflows would be better as a single prompt. The hierarchy goes all the way down.

What they're actually saying, if you squint, is that the complexity lives in the problem — not in the infrastructure you pile on top of it. A "prompt chaining" workflow is four lines of Python with two API calls. It does not require a framework with a YAML config file and a visual debugger. It requires you to understand why you're chaining the prompts.

The framework installs the understanding for you, which is exactly the problem.

Every assumption a framework makes correctly is one fewer thing you had to think through. Every assumption it makes incorrectly is a bug you cannot diagnose because you don't know what the framework assumed. This is not new — it's the same reason you eventually had to read how React actually works even though you were productive with it before you did — but it's particularly sharp here because agent behavior under load is genuinely weird and the last place you want mystery is in the layer that's making decisions.

Start with the API. Add complexity when the absence of it becomes the problem.