Amey-Thakur Generate and verify
A single agent asked to produce and check its own work grades itself generously, because the reasoning that produced the error also evaluates it. Splitting generation from verification, with the verifier told to refute rather than review, is the highest-value multi-agent pattern there is.
Method
- Give the verifier the opposite goal. Not is this good but find what is wrong with this, defaulting to rejection when uncertain. A verifier asked to review agrees.
- Withhold the generator's reasoning from the verifier. Give it the output and the source material, not the chain that produced it, or it inherits the same mistake (see agent-context-isolation).
- Require evidence for both verdicts. A rejection names the specific flaw; an acceptance names what was checked. Unexplained verdicts are noise.
- Use several verifiers with different lenses when it matters. Correctness, security, and does-this-reproduce catch different failure classes, and diversity beats repetition (see agent-ensemble-voting).
- Return failures to the generator with the specific objection. Regenerating from scratch loses the good parts; targeted revision keeps them (see agent-iterative-refinement).
- Cap the loop. Two or three rounds, then escalate to a human, because a claim that survives no verification round will not survive the fourth either.
- Track which verifications actually catch things. A verifier that never rejects is either unnecessary or not doing its job.
Boundaries
Verification raises confidence; it does not prove correctness, and a verifier sharing the generator's blind spot confirms the error confidently. Cost roughly doubles, so apply it where consequence justifies it. For anything irreversible, human review remains the last gate (see agent-human-checkpoint).