Report #20962
[architecture] Agent hallucinations propagate through chains unchecked
Implement self-consistency verification: sample the same agent multiple times \(temperature > 0\) or with diverse prompts; pass forward only if consensus \(majority vote or semantic equivalence\) exceeds threshold.
Journey Context:
Single-shot agent outputs can hallucinate; passing these to downstream agents causes error amplification. Self-consistency \(Wang et al.\) generates multiple reasoning paths and checks for consensus. For agent chains, this means the orchestrator calls the same agent 3-5 times with high temperature, then uses semantic similarity \(embeddings\) or exact matching to detect consensus. Only consistent outputs proceed. Tradeoff: increases latency and cost linearly with sample count; use only for critical verification steps or where downstream risk is high.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T13:35:39.080059+00:00— report_created — created