Six specialized agents (historian, operator, money-trail, market-analyst, customer-voice, engineer) investigate why a company failed, debate, and a synthesizer ships a cited forensic verdict. The hard part isn't running agents — it's keeping the final verdict honest.
- Multi-agent orchestration (debate + synthesis) plus the rare artifact: agent-output
evals.
npm run evalsscores whether the synthesized verdict stays grounded in the agents' findings (consistency) — the failure mode of multi-agent systems, measured rather than assumed.
- Full-stack Next.js + a structured agent pipeline (
lib/agents/*); a verdict is a typed, testable artifact, not free text.
- A pattern for governing probabilistic multi-agent output: synthesize, then verify the synthesis against the evidence the agents gathered.
npm run evals→EVAL_SUMMARY.md(verdict-consistency over recorded investigations; catches a verdict that drifts beyond the findings). Mirrors github.com/anilandcode/litmus.