revue de code

Brian Houck's figures set the terms: at Meta, significant lines per human-merged diff rose 106% in a year and diffs per developer per month 51%, more than 80% of that growth attributed to agentic AI; DX measured median pull request size up 64%. Developers want to spend about 7% of their time reviewing.

The disagreement turns on a 2013 finding that Houck and Martin Monperrus both cite: in Bacchelli and Bird's Modern Code Review study, defect-related comments were 14% of the comments actually written, while defect detection was the most frequently stated motivation. More than half of developers said they reviewed to explore alternative solutions, transfer knowledge, and see what teammates were building.

Monperrus, in a June 2026 arXiv position paper, treats those objectives as reachable by autonomous LLM systems at lower cost and higher throughput, names inspection de Fagan (1976) as his historical target, and proposes a multi-agent adversarial verification pipeline with a consensus vote. Rachel Laycock, CTO of Thoughtworks, keeps every function and moves it earlier: alternatives explored before one is implemented, knowledge transferred by pairing, collective ownership from mob programming and whiteboard design, architectural intent carried by shared design and fitness functions. Review by exception is what survives.

Max Kanat-Alexander, writing from Capital One in November 2025, stays with the load: AI turns every developer into a full-time reviewer, and rubber stamping lets mediocre code through. Houck and Laycock both accept Margaret-Anne Storey's cognitive and intent debt as real. They split on whether the mandatory pull request is any defense against it.