Adversarial Verify

Eliminate false positives with competing agents.

The problem: Agents are sycophants. Ask them to "find bugs" and they will find bugs whether they exist or not.

The solution: Three agents compete.

1. Bug-finder gets points for bugs found (superset of possible issues)
2. Adversarial agent gets points for disproving bugs (subset of actual issues)
3. Referee scores both and outputs verified list

Commands

/verify [code-path] - Run adversarial verification /verify --focus security - Security-focused review /verify --focus performance - Performance-focused review

Attribution

Pattern from @systematicls (World-Class Agentic Engineer)