Dave Currie

AI Janitor: Making Architecture Review Executable for Coding Agents

AI coding agents accelerate code generation, but they also accelerate architectural drift, regressions and false confidence. While building a multi-service production platform with only two of us, I found that a normal human review loop could not keep up with the volume of AI-generated change. I needed some of the boundaries and feedback a platform team would normally provide, so I began turning repeated review comments, architectural decisions and past failures into checks that agents could act on themselves.

This is a practical case study focused on two mechanisms. First, custom ESLint rules whose messages tell an agent what went wrong, why it matters and how to fix it, giving the agent enough information to correct its own change before human review. Second, fail-closed verification: what I learned from a guardrail that stayed green while silently doing nothing, and how I changed the system to prove that its checks actually ran.

I’ll show where these controls work, where they become lint theatre and which decisions still need a human. The point is not to encode every architectural decision or replace review. It is to make the repetitive, objective parts executable and reserve human attention for judgement. Attendees will leave with a method for identifying repetitive review work and turning the right parts into fast, mechanical feedback—without building a huge internal platform or trusting AI output by default.