Can one consequential AI-agent execution be reconstructed end to end? Use this checklist to test whether the evidence chain survives after the workflow runs.
Identity
- Which human, service account, workload identity, or delegated agent initiated the execution?
- Is there a durable identifier that can be correlated across systems?
Authority
- What permission, scope, delegation, or role authorized the action?
- Can you prove the authority state that existed at execution time?
Policy
- Which guardrail or policy was evaluated?
- Is the runtime policy decision retained?
Action
- Which tool, API, data source, or system was touched?
- Are consequential parameters and timestamps retained?
Approval
- Was human approval required?
- If yes, who approved, what was presented, and what exactly was authorized?
State change
- What changed in the target system?
- Can the resulting state be linked back to the execution?
Evidence integrity
- Can records be joined across identity, policy, action, approval, and resulting state?
- Are there tamper-evident or independently verifiable signals where needed?
Decision gate
If one or more layers cannot be reconstructed with confidence, the workflow has an evidence gap worth resolving before broader scale.
The 48-Hour Agent Execution Assurance Sprint applies this model to one bounded workflow for a fixed $1,250 scope.