Multi-agent "shared brief" built from precomputed data goes stale when agents plan on a what-if state
Pattern: a deterministic situation brief (one page of facts injected into every LLM agent prompt) was built from a precomputed replay of the recorded hour, while the agents physics tools simulated the live state. After a what-if mode was added (scaled weather, lines already switched off), runs on that state got a brief describing the recorded hour: humidity +100% gave 40 lines over the alarm in the brief versus 0 in physics, and a line already switched off was still listed as a shut-off candidate. Agents would then reason from contradictory facts. Fix: (1) add a state.is_modified() (non-zero conditions or any operator changes); (2) when modified, take the brief fields that depend on the inputs (risky lines, region index, weather extremes) from the state own simulation, via one shared helper also used by the UI simulate endpoint, so the brief and the map numbers cannot drift apart; (3) keep fields that cannot be recomputed for one hour (a time-integrated storm index) or that are real external facts (official warnings) from the record, and document that; (4) include the what-if inputs in the brief cache key; (5) add a regression test asserting brief counts equal physics counts on a modified state. Also useful: build the deepagents graph offline with a dummy API key to verify every tool name resolves and every prompt carries the brief, without network or spend.