Context

Changes large enough that a human review is the bottleneck, and small enough that a second pass can actually read them. Also useful as a filter: most diffs get one round of critique before a human sees them at all.

Problem

The agent that wrote the code cannot review it. Not because of capability - because its context is contaminated: it remembers its intentions, it believes its choices were justified, and it will read the diff as evidence for the plan rather than as an object to critique. Ask an author to review their own work and you get a summary of the plan, restated.

Pattern

Spawn a reviewer with:

  • Fresh context. It sees the task statement and the diff. Not the author's transcript.
  • The specification, not the summary. Point it at the acceptance test, the issue, the invariant.
  • A required output shape. Findings, each with file, line, severity, and what would have to be true for it to be a real problem. Unstructured review output is unactionable.
  • An explicit mandate to find at least one thing - otherwise the default answer is "looks good", which is exactly the failure mode you are trying to avoid.

The parent then triages: apply the real findings, discard the noise, and re-run the reviewer once. Two rounds is usually the point of diminishing returns; beyond that, the reviewer starts inventing issues to justify its existence.

Tradeoffs

Costs roughly half again as much as a single run, and produces false positives that a human must sift. A reviewer that is systematically pessimistic trains you to ignore it.

When not to use it

Trivial diffs, and anything where the bottleneck is the human understanding the change rather than spotting errors - in that case the reviewer adds a layer of summary between you and the code, which is worse than reading it yourself.