GUS: company research agents
I built a due-diligence tool where 8 AI agents research a company in parallel, covering financials, legal record, leadership and risk, and anything without a source is left out of the report.
01The problem
AI can write company research quickly, but you can't tell which claims are real. In due diligence, a confident claim with no source is worse than no claim at all.
02What I did
I designed and built the part that coordinates everything: the planner, the agents running in parallel, the tracking that ties every claim to its source, and the final check on what goes into the report.
03What made it hard
A made-up claim costs far more than a missing one. So the tool has to prefer saying nothing over saying something that only sounds right, which is the opposite of how these models normally behave.
04How it works
A planner breaks the question into parts and sends each to a specialist agent. Every finding keeps a link to its source. The final check drops anything without one, and the report shows that it was dropped.
05Other options I ruled out
One big agent would have been simpler, but then it's hard to tell where each claim came from. Fact-checking a finished report was also too late, because by then other parts of the report depend on the unsupported claims.
06How I tested it
I asked it questions whose answers were not in the sources, to check it left a gap instead of inventing something.
07Results
The repo is public: the planner, the agents, the final check, the dashboard and 45 tests that run in CI.
08What's next
A public benchmark with full run logs, so anyone can check the results.