The slice from a few days ago proved one agent could reason over shared evidence and produce cited findings. The obvious next move is "add more agents." The real work is everything around that.
Three things had to be true before a second agent was worth adding:
One shared source of truth. Every agent reads from the same evidence store, scoped to the same engagement, and writes findings back into the same place — each still tied to the evidence it came from. No agent gets its own private view of reality. Cross-engagement leakage stays impossible; that boundary is enforced, not hoped for.
A stable contract. Each agent takes the same shape of input (scoped evidence) and returns the same shape of output (findings, with citations, plus a confidence signal). That's it. Adding a new kind of analysis is now a small, well-bounded change instead of a rewrite — which is the whole point of doing the plumbing before the second agent, not after the fifth.
Failing loudly. The orchestrator runs multiple agents over one body of evidence, in parallel where it's safe. The interesting case is when one of them falls over. The rule I'm holding: a single agent failing must fail loudly — it never corrupts the run, and it never silently disappears while the others quietly carry on as if the picture were complete. A partial answer that looks complete is worse than an obvious error. There's a test for exactly that: kill one agent mid-run, assert the run surfaces the failure and preserves everyone else's work.
The theme underneath all of it: in a system whose entire value is trustworthy output, the coordination layer is not glue code. It's product. Every agent action is recorded and reconstructable — who ran, over what evidence, when, and what they concluded.
More as it ships. I'm logging the whole build in my Lab.