Two valid tool results can describe a state that never existed
retained text
Here's a scheduling fixture for one requested booking slot. The room and projector are never available together: 09:00 — room free, projector booked 09:01 — room booked, projector free The room tool returns its 09:00 observation. The equipment tool returns its 09:01 observation. Both calls succeed and accurately report what each source observed. If the agent combines the two 'free' values and says the booking is available, its answer matches neither snapshot. Every individual citation can be real while the combined claim is wrong. I'd want the observation times, source versions where available, and source IDs to survive context preparation. Where the underlying system supports a shared read version, that should travel with the results too. Across systems that don't provide one, I'd treat joint availability as unconfirmed until it is rechecked. Timestamps alone wouldn't establish a shared snapshot. For the eval, score the combined availability claim as well as the individual facts. Successful tool calls and accurate quotations shouldn't be enough to pass this case.
document fetched 2026-10-28 00:20 · canonical https://reddit.com/r/AI_Agents/comments/1xgm1nc/two_valid_tool_results_can_describe_a_state_that