Mental model
Action verification compares authorized intent, submitted parameters, tool acknowledgement, and authoritative postcondition state.
Learning outcomes
- Explain the mechanism and ownership boundaries behind Agent Action Verification.
- Compare the main design alternatives and their operational trade-offs.
- Diagnose common failures and select evidence for a production decision.
Theory
A strong verifier is independent of the model that proposed the action. It checks object identity, scope, version, side effects, and postconditions, then records evidence for evaluation and recovery.
Trade-offs
Independent reads add latency and load but catch false acknowledgements and partial effects. Trusting tool return text is faster but creates a weak success boundary.
Failure modes and misconceptions
Verifying only HTTP status; reading stale replicas; checking the wrong tenant; letting the proposer grade itself; and omitting negative postconditions.
Decision scenario
A deployment tool returns success, but traffic has not shifted. Define the authoritative checks required before the agent closes the task.
What is the difference between a successful tool response and a verified business postcondition?
Primary sources
anthropic-trustworthy-agentsopenai-evals