François Chollet, Keras creator, warns that agentic coding triggers a feedback loop compounding codebase complexity exponentially

@fchollet Just wrote about how complexity accrues in the prompt:

@fchollet adding this to my AGENTS.md. that'll fix it! 😏

With agentic coding, complexity compounds in a mechanical way: unnecessary code ends up in the codebase, moves to the context window, degrades the model's reasoning abilities, leads to more unnecessary code (often to fix issues arising from the unnecessary code). It's exponential

@fchollet True! This is why a solid mental model with a good dev in the loop is necessary. IMHO vibe coding will do a lot of damage to good software engineering

@fchollet that’s why I built Lore, so you can capture decisions and reasons - give your agents the WHY - so it doesn’t drift off into oblivion

https://github.com/itsthelore/rac-core

@fchollet It is astonishing how bad most programmers are at looking into where this tech is going. You are just focusing on issues right now, and tomorrow you focus on tomorrows issues. All these problems will fall as the tech matures.

@fchollet Do you think compounding complexity is intrinsic to agentic coding or just the current behavior?

@fchollet This is why I keep giving agents smaller jobs. The bigger the task, the easier it is for them to make a beautiful mess.

@fchollet I've been likening it to the elephant toothpaste chemical reaction. Infinitely expanding code.

@fchollet I'm designing an agentic workflow to fix this, it goes from vibe coding -> TDD using spec driven development. I'm trying to combine the best of all current methods

@fchollet @sean_a_mcclure

@fchollet funny part is watching an agent write code to fix the code it wrote yesterday. compounding works both ways. context rot is just tech debt billable by the token.

@fchollet Agentic coding doesn’t just create debt it creates context pollution that makes models progressively dumber.

@fchollet Unnecessary code only compounds that fast when the loop has no explicit pruning or summarization step before the next context window. Most setups ship the first version of that loop and call it done.

@fchollet This has always been the case

@fchollet Context windows are finite. every line of bloat shrinks the model's intelligence. agentic coding is just auto-inflating the cost of reasoning.

@markerdmann @fchollet Would you really see adding this in .md file would fix it.

I have seen that these .md files needs to hold as much dense information as possible such that the agent will know the boundary but still human at POC should know what the code does in the repo else it is a mess.

@A_K_Nain @fchollet I have seen that, by definition(going forward), a good dev should know what the code does in the repo; there is a possibility of slap code getting added, and the dev didn't notice it. I have seen this firsthand

@fchollet messy code spirals when ai writes without guardrails. a single strict agent stops the chaos before it grows

@fchollet applies double with agentic coding. the context window is a codebase - prune aggressively or complexity compounds mechanically.