Two approvals are not two observations.

Two approvals are not two observations.

Short video caption

Two reviewers can agree from one incomplete context. Give the reviewer the original contract and independent evidence, demand a counterexample, then run a falsifying test. Illustrative migration. #EricFieldNotes

Full written post / accessible read

Agent-to-agent review can catch many errors. It can also multiply confidence in one shared assumption if both reviewers receive the same incomplete ticket, tests and summary.

In an illustrative migration, the author and reviewer both see an average-load benchmark. Neither sees the tenant's peak-lag limit. Both approve a plan that would have failed the real acceptance condition.

Give the reviewer the original contract and incident history, not merely the implementer's summary. Ask it for a counterexample and a predicted test result. Then run the test in a setting that can overturn both recommendations.

Treat model agreement as a proposal until it survives a test neither agent controlled. Save the source, divergent hypothesis and result with a human owner. Do this because independent evidence, not the number of voices, reduces decision risk.

#EricFieldNotes

Evidence and boundary

On-screen boundary: ILLUSTRATIVE MULTIAGENT REVIEW. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.

• Huang et al., Large Language Models Cannot Self-Correct Reasoning Yet (S76)
• Zhu et al., Demystifying Multi-Agent Debate (S77)
• Ki et al., Multiple LLM Agents Debate for Equitable Cultural Alignment (S78)
• Liu et al., Lost in the Middle (S80)