Adversarial Consensus Line

The hardest part of a multi-agent factory isn't getting work done — it's what happens when two competent agents disagree. This blueprint runs a deliberate adversarial pair: Solver A stays conservative, Solver B explores. A consensus node scores both proposals against the acceptance criteria and, on a genuine conflict, re-opens a bounded debate instead of silently picking one. The loop closes only when the verifier signs off, so no human ever has to arbitrate.

Use it wherever a single model is confidently wrong too often: spec interpretation, refactor strategy, or any judgement call where a second, differently-tuned opinion catches the blind spot.