AI Coding Agents Are Breaking Diff-Based Code Review

Code review was always a compromise: reviewing the diff instead of directly verifying intent, because the diff was the only artifact that was legible, diffable, and line-commentable. That worked when a human wrote every line, since variable names, structure, and reasoning left traces a senior reviewer could reconstruct.

Agentic coding breaks that compromise. An AI-generated diff isn't an approximation of intent — it's the rendering of a prompt, a test suite, and mock or contract constraints into code. There's no reasoning trail in the diff to infer from, because the reasoning never happened there. A diff can pass CI, look clean, and still ship the wrong specification.

Intent hasn't disappeared; it has moved to three under-reviewed surfaces: test names (the scenario index), builder and factory APIs (the domain vocabulary), and interaction or contract expectations (the system's boundaries). Teams that keep reviewing only the diff will see review quality hold steady while what actually gets built quietly drifts — missing scenarios, inconsistent vocabulary, unchecked boundaries — none of which shows up as a red flag in the pull request.

For engineering teams adopting AI coding agents, this reframes what 'reviewing a PR' should mean: start upstream, at test names and vocabulary, before ever reading the implementation diff.