Second opinion

You are orchestrating a review where the reviewer must be a DIFFERENT model than the author of the change. A model reviewing its own output shares its own blind spots. A second, independent model has different ones.

Step 1, Collect the change

Gather the diff under review (staged changes, the branch against its base, or the named PR) plus just enough surrounding context to judge it. Note which model authored it, if known.

Step 2, Brief the independent reviewer

Send the diff to a different model, different vendor if available, a different model at minimum (a second CLI, an API call, another agent). The briefing stance, verbatim:

This code was written by another AI. Your job is to prove it wrong: find the bug, the missed edge case, the unsafe assumption, the quiet behavior change. Judge the "why", not just the "what". If you cannot fault it, say so plainly.

Give the independent reviewer the diff and the context. Do NOT give it the author's reasoning or self-review, that anchors it.

Step 3, Triage the verdict

• Clear, the independent reviewer raises no substantive objection: record one line ("second-opinion: clear") and pass.
• Objection, give the author model ONE rebuttal round with the objection verbatim. If the author concedes, apply the fix. If it stands its ground, escalate.

Step 4, Escalate disagreements to the human referee

Present both positions verbatim with file and line references, plus your own one-line read. NEVER resolve a standing disagreement silently, the disagreement queue is the highest-signal reading the human referee gets.

Rules

• The independent reviewer never edits code. It only argues.
• One review pass, one rebuttal. More rounds means the models negotiate; you don't want them to.
• A wrong "why" is a real objection even when the code happens to work.

---

By @runsagents, agents write the code, second, independent models check it, and human referees read the arguments.