An AI agent completes its task in staging. But can it access another customer’s records, follow instructions hidden in a retrieved document, or take an action without approval? Before shipping, define which failures should stop the release. For a support agent, that might mean: Accessing records outside the signed-in customer’s account. Following instructions in retrieved content that lead to unauthorized behavior. Issuing a refund without the required approval. Test those boundaries against the connected workflow using controlled data. Check the outcome as well as the response: “I didn’t issue a refund” means little if the backend created one. When you can’t verify a critical action, mark it for review. Keep the evidence, fix reproducible failures, and rerun those cases with the next candidate. That’s the release checkpoint we’re building AI Release Assurance around: helping teams turn observed behavior into a decision to approve, block, or review a release. Read the full ADLC article on PromptBrake, or follow AI Release Assurance on LinkedIn for more practical notes. What failure would stop your team from shipping an agent?