revue de code agentique

Anthropic's monorepo went from roughly 2,500 daily changes in May to more than 30,000 in September. The Claude Code team, reviewing code about 90% of which Claude wrote, dropped line-by-line human reading and swapped the question "has a human read this?" for "which control should catch what can go wrong?"

What replaces it is layered. Before the pull request, Claude runs tests, build and linter, then a subagent re-reads the diff in a fresh context. CLAUDE.md describes the project; REVIEW.md, written by the code owners, tells the reviewer what matters. Several agents hunt bugs in parallel, others vote on each finding, and only retained findings get published. A separate /security-review and CI block the merge. A classifier sets the risk level, which fixes the number of human approvers at zero, one or two. An agent's approval never counts as a human's.

Martin Monperrus pushes the logic further in a June 2026 arXiv position paper: human review is no longer a necessary component of a quality pipeline. He aims at Fagan inspection (1976), cites Bacchelli & Bird on review catching fewer bugs than developers assume, and reads the arrangement where an agent writes and a human approves as false security. He ties review throughput to DORA metrics.

The friction is agent noise. One pull request chained more than fifty review rounds in a night, so the Claude Code team capped rounds at three and limited re-review to the new diff, moving the average from 2.08 to 1.32.