Guardrails in practiceยท

Notes on reviewing AI-written code

An agent doesn't learn from a code review, AI code looks finished before it is, and approval prompts stop being read.

An agent doesn't learn from the review

When a junior developer gets feedback in a code review, they learn from it. Their next pull request is better.

An AI agent doesn't. Every session starts fresh, so the same mistake comes back next week.

Unless someone writes the lesson down where the agent reads it every time, like a CLAUDE.md or AGENTS.md file, the review never gets any cheaper.

Review time belongs in the estimate

AI code arrives so fast it looks finished.

That's why review time gets squeezed at the end of the job. It needs to go in the estimate from the start, same as testing.

Approval prompts stop being read

If an AI agent asks for approval ten times a day and nine are fine, people stop reading and just approve.

The approval step is still there on paper. Nobody is really checking.

First posted as three notes on Substack between 26 and 28 October 2026.

More in Guardrails in practiceAll articles