Cross-agent second opinion for AI output verification

★ 6.5 · general

cross-agent-second-opinion routes a first agent's output in Claude Code to a second independent agent for verification — not revision, not decision-making, but surfacing risk, inconsistency, and missing evidence before a human acts on the result. It is designed for high-stakes outputs where a single-agent error could alter a scientific conclusion, corrupt a citation set, or introduce a defect into a release package, and for cases where two agents have already produced conflicting outputs that require human adjudication. The second agent receives only the files scoped for the review; sensitive data is not duplicated without prior anonymization. Every disagreement is classified into one of four categories — human decision, source verification, code execution, or data boundary — and any agreed corrections are held as small traceable diffs pending explicit approval. Agreement between two agents is explicitly not treated as proof of correctness, since both can share the same blind spot.