Run receipts, evaluation, and incidents
Record attributable action traces and test success, policy compliance, recovery, and incident response.
Run receipt
A run receipt is a structured, integrity-protected account of what an agent was authorized to do and what occurred. It should identify actor, agent and policy versions, purpose, input references, plan or decision points, tool calls, arguments or protected references, approvals, results, side effects, errors, retries, budgets, final state, timestamps, and evidence links.
The receipt should distinguish proposed, approved, attempted, succeeded, failed, compensated, and uncertain actions.
Evaluation
Test task success together with permission, tool selection, argument validity, approval adherence, confinement, sensitive-data handling, side effects, stopping, recovery, and receipt completeness. Include prompt injection, malicious tool output, stale state, conflicting instructions, partial failure, repeated retry, unavailable approver, and revoked permission.
Incidents and recovery
Prepare to halt runs, revoke credentials and tools, isolate workspaces, identify affected resources, preserve evidence, compensate reversible actions, escalate irreversible effects, notify owners, and add regression cases.
Do not let missing telemetry silently permit continued action. Define whether the agent pauses, stops, or operates in a separately approved restricted mode.