Define the promise
Start with an explicit contract: the jobs an agent can perform, the actions it must refuse, and the evidence required before it can claim a state change.
A useful contract turns vague requirements such as safe support into testable behavior such as refusing credentials, escalating unsupported account changes, and avoiding unverified refund claims.
Test a live endpoint
Verification must reach the same public HTTPS endpoint used by the deployed agent. AgentProof sends scenario inputs, records bounded response evidence, and evaluates happy, boundary, edge, and adversarial behavior.
Share the outcome
A reliability score is only useful when the underlying findings are inspectable. Private reports retain full evidence for the owner, while a public report can communicate an unexpired verified result without exposing credentials or account data.