Using Linters to Direct Agents

Interesting read that resonates with what I have been doing lately.

I am really glad to see good engineering practices such as linting, proper code architecture, and no network calls in tests, etc. are now paying off by letting agents move faster and more precisely through more automated self-healing loops and more guardrails.

It is also interesting to see that some things that were initially done to simplify human readability of code, such as default imports, may now have a negative impact. I wonder if this could also extend other things such as adding explicit type annotations everywhere instead of relying on implicit ones, or using full variable names instead of `x`, `i` or abbreviations, which are all currently written to optimize human consumption.

I am also curious to learn what the sweet spot is between enforcing rules for humans versus code written for agents. My gut feeling is that it is now time to start optimizing for agents.

I wrote up how we use linters as the “law” for AI coding agents. The gist: encode expectations as lint rules (searchability, folder layout, architecture boundaries, tests, security) so “lint green” = done, cutting the human-in-the-loop bottleneck and making mass refactors reliable.