Single-agent validation of complex outputs (factual accuracy, code correctness, risk assessment) is unreliable because the validator shares the same biases and blind spots as the generator.
Solution
Route critical outputs through a quorum of N independent validator agents using different prompting strategies; require M-of-N agreement to pass — reject or escalate on disagreement.
Implementation steps
Define quorum size N and threshold M (recommended: N=3, M=2 for speed; N=5, M=3 for high-stakes)
Spawn N validator agents independently — do not share intermediate results between validators
Each validator applies different evaluation lens: factual accuracy, logical consistency, completeness
Collect verdicts: PASS, FAIL, or UNCERTAIN with required reasoning
If M or more PASSes: accept output. If M or more FAILs: reject with aggregated failure reasons. Otherwise: escalate to human or senior agent.
Examples
–Medical diagnosis validation
–Financial risk assessment sign-off
–Code security review
–Legal clause analysis
Anti-patterns
–Using same base prompt for all validators — introduces correlated failures
–Treating UNCERTAIN as PASS — uncertainty requires resolution, not assumption
–Skipping quorum for "routine" outputs — selection bias creates false confidence
