Recorded operations dashboard
A real session graph, captured end-to-end from a live google.adk fleet run on GKE — Vertex Gemini agents reasoning over MongoDB Atlas, Snowflake, and BigQuery. Every number below is read back from the runtime: the agents and their scopes, the append-only belief log, the contradiction that genuinely fired and how trust resolved it, the federated fetches, and a counterfactual branch. Nothing here is staged.
Agent fleet
6 spawned- race_strategy strategy, tyres committed
- tyre_engineer tyres committed
- aero_rd aero no commit
- logistics parts, supplier, calendar committed
- telemetry_analyst telemetry committed
- weather_model weather committed
Three-tier memory
hot · warm · coldBelief state & reconciliation
contradiction resolved- parts All necessary parts for the final stint are confirmed to be in optimal condition and readily available. logistics
- telemetry Current car telemetry data is unavailable. telemetry_analyst
- tyres Final stint: softs. race_strategy superseded
- tyres Final stint: hards. tyre_engineer accepted
- weather There is a 30% chance of light rain in approximately 15 minutes. weather_model
Federated fetches
live · provenance- aero_rd bigquery:ezra-498021.formula_1.aero_configs 325 rows time-travel
- race_strategy snowflake:EZRA.PUBLIC.RACE_RESULTS 25873 rows time-travel
Permission scoping
policy engineThe policy engine refused the out-of-scope fetch before any data left the source — scoping is enforced by Ezra, not the agent.
Branching replay
counterfactualA real counterfactual branched from turn 1 of the live graph and run forward — mutating race_strategy's call to "wets at lap 43". The diff shows where the branch diverges from what actually happened.
Recorded from a live GKE run, inherited from race-weekend-imola-2026. The same
state is available live over the REST surface (/ezra/health, /ezra/belief/snapshot, /ezra/branch/diff). See the REST API.