FDE01Forward deployment
Handbook 08

Verify what happened when AI says it is done

Handle success, failure and unknown outcomes separately.

On this pageChapter guideDiscussion questionRelated exerciseReferences

Chapter guide

Inspect identity checks, data, retrieval, rules, the model, tools and authoritative business records separately. Locate the failing step before choosing a repair; not every failure is a model hallucination.

A tool timeout can mean the action happened or did not happen. Check the business record before retrying. Idempotency means repeating the same request does not create another business action; it must be implemented and tested, not assumed.

For long-running agents, inspect session records, the execution loop and the runtime. Check persistence, recovery, permissions and isolation. Managed infrastructure still leaves customer-side data and business responsibilities to assign.

Where model output belongs in a workflow
Where model output belongs in a workflowIn this example, an authorized person reviews the output before the business record is updated. Reviewers need evidence, time, skills and the right to reject.Task & permissionIdentity / evidenceModel & toolsAdvice / proposed actionHuman reviewApprove / reject / take overSystem of recordAuditable final stateMissing evidence or authority: pause, investigate or hand off
In this example, an authorized person reviews the output before the business record is updated. Reviewers need evidence, time, skills and the right to reject.Swipe to view the full diagram.

Discussion question

Where do you confirm that an action happened when the model says it did?

Related exercise

Follow the related lesson below and try its exercise. Use the linked template to record your judgment, evidence and open questions, then revise with the checklist. With a partner, check which parts of each other’s work need clarification.

References

Anthropic / Scaling Managed Agents: Decoupling the brain from the hands