Transcription
I'm gonna tell you something crazy. 98% of Clawude code is not AI. And I know that sounds crazy, but Enthropic never gave any details about how it worked. There's no docs or architecture, just a tool that basically every dev is currently obsessing over until it was leaked a few weeks ago where a research team pulled the source code and went through all 500,000 lines of code.
And what they found is that the actual AI, the part making the decisions, is 1.6% of the whole thing. It's basically a while loop with one function called call model. It gets handed a bunch of contacts, makes a decision, calls a few tools, and passes the output back. That's basically the whole AI part. Everything else is the harness catching that request. For example, there's a permission system with seven different modes checking if it's allowed. A five layer pipeline compressing your conversation history so the AI doesn't lose context mid task. 54 tools doing the actual execution. There's also a big focus on recovery systems for when things break.
So, the AI is basically a consultant sitting in a room. It only speaks when asked and an entire operation runs around it, making sure what it says actually happens correctly. And it's a shame because so many companies right now are racing to make AI agents more autonomous, more handsoff, and less humans in the loop. But what this paper shows is that the thing making Claude code work successfully isn't the AI. It's the 98% wrapped around it. The harness, including the guardrails, the recovery systems, permissions, that's the thing that drives everything together.
The paper even says as models keep improving, they're all going to start performing similarly. So, what separates the winners? It's the infrastructure built around it, not the model. Which means the whole just let AI do everything approach won't get you too far. And it sucks because I don't think the people building these things want to hear that. follow and I'll keep you posted.