Agent runs and versions

Every agent run should identify the agent version, input, context, model route, tool calls, policy decisions, outputs, cost, and final status. A version is the unit you test and promote; a run is the evidence of what happened.

Build a case set from real work, compare the candidate with the production version, inspect any regression, and record why a promotion was approved. If production behaviour degrades, promote the previous version and preserve the affected traces for investigation.