Uncle Bob says he no longer reads AI-generated code, relying on tests and QA instead

机器之心 · wechat · 2026-07-24

Uncle Bob says he no longer reads AI-generated code line by line

A Machine Heart article summarizes a debate sparked on X: once AI agents can break down tasks, edit files, write tests, run code, and iterate on errors, should developers still review every line they generate?

Robert C. Martin (“Uncle Bob”) argues for a more radical workflow:

He says he does not read agent-written code at all.
Instead, he surrounds the agent with strict guardrails: unit tests, Gherkin tests, QA flows, quality metrics, mutation testing, coverage targets, and other automated checks.
In his view, this shifts quality control from manual code reading to an industrial-style verification system.

The article also lays out the counterargument:

Many developers feel they must understand any code they are responsible for.
Tests and constraints help, but they do not guarantee the agent did not miss requirements, change tests to fit its own implementation, or leave deeper bugs undetected.
Even in an AI coding workflow, many developers believe some level of code review remains necessary.

The practical question is no longer whether to review code at all, but which parts must still be reviewed, and how much can be delegated to automation.

Original post →

Uncle Bob says he no longer reads AI-generated code, relying on tests and QA instead

More from coding & agent