Multi-Agent Verification for Reliable AI Research

AI agents are confidently wrong. So I made mine check each other's work.

Last week I gave 13 agents one job: research crop-disease apps for smallholder farmers. 39 sources, 16 findings.

Before anything reached the final report: → every claim was traced to a page the agent actually opened → a verifier re-read the riskiest pages against the real text → 3 independent passes voted, and only repeat findings survived → a critic tried to prove each gap was already solved

What came out: 5 unsolved opportunities and one concrete experiment to run first.

It all runs on free-tier AI, and it's open source: github.com/Samthnalyst1234/Farmbt

Question for the automation folks here: how do you stop your agents from making things up? Human review, a second model, or something else? 👇