note · 2026-09-15

Code review of AI written code, review by exception instead of reading every line

When agents write most of the code, requiring a senior to read every change turns the seniors into the bottleneck. Rachel Laycock, CTO of Thoughtworks, argues for moving knowledge sharing, design and deterministic checks earlier, and reserving human review for a short written list of exceptions.

For CTOs and engineering leads whose review queue grows faster than their outputRoc Rizzardini, CEO of yaab1 min read
Code review of AI written code, review by exception instead of reading every line

If every line an AI agent writes has to wait for a senior engineer to read it, you have not built a faster team. You have built a queue.

Rachel Laycock, CTO of Thoughtworks, made that argument on martinfowler.com on September 2, 2026, answering a colleague from DX who wants to keep human review of everything. She cites his numbers: at one large company, lines of code per human landed change up 106% in a year; across DX's data, median pull request size up 64%. Her position is that code review has been carrying too many jobs and that most of them belong earlier: pairing, design sessions, constraints encoded as tests, anything deterministic automated. Humans review by exception: an architectural change, a sensitive security boundary, a large blast radius, unfamiliar critical code, or anyone saying "I am not confident about this."

That is the rule I run inside client teams. The exceptions are written down. Everything else passes automated checks and an agent review against a rubric before a person is assigned.

Kent Beck, creator of test driven development, quoted by The Pragmatic Engineer on X: feedback comes in layers like a filter; the bad stuff sticks, the good stuff still goes through. Review is one layer. It was never meant to be the only one.

Sources