Automated review

Automated review is a model reviewing the work of a model. A separate session, often a subagent or a bot on the pull request, reads the diff with fresh context and reports what looks wrong: a missed edge case, a security smell, a change that doesn't match the spec. It sits between an automated check, which is exact but narrow, and human review, which is expensive but has taste.

The reason to use a fresh reviewer rather than asking the same session to check itself is sycophancy plus contamination. The session that wrote the code carries every assumption that produced the bug, so it tends to confirm its own work. A reviewer that only sees the diff and the requirements has none of that baggage and catches things the author can't.

Treat its output as a triage list, not a verdict. Point the reviewer at specific concerns (correctness, security, the acceptance criteria in the ticket) instead of asking for a general opinion, and have it verify each finding by reading the surrounding code before reporting. Anything flagged with evidence goes back to the author session as a fix; anything vague gets ignored. It reduces what a human has to read; it doesn't replace the human.