AI agent acceptance checks: prove the task finished

What to inspect

Define the observable outcome outside the agent transcript. Check the saved artifact or permitted state change, then decide where a failed check goes. A run status can describe execution without establishing that the requested work is useful.

Worked example

For a report task, check that the file exists, opens and contains the required sections. For a record update, check the allowed fields on the intended record. Keep a failure owner and a recovery step in the same worksheet.

Copy the full prompt

Adapt the inputs to your task. Remove private data before sending anything to a model. This page copies text locally; it does not run the prompt.

Turn [WORKFLOW] into an acceptance contract. Include expected artifact or state, a deterministic validation step where possible, allowed side effects, timeout, failure owner and recovery action. Identify what can be verified automatically and what needs human review. A completed run status alone must not count as acceptance. Return a table with blank fields clearly marked.

Review checklist

1. Define the expected artifact or state.
2. Verify the result independently of the status message.
3. Name a failure owner and recovery action.
4. Record what still needs human judgment.

Keep the resource

Keep exploring

• LLM evaluation: measure the usable result
• AI evaluation cases: turn a failure into a regression test

Documentation & scope

Anthropic: building effective agents ↗

This is an original workflow template with an illustrative example. It does not report completed model tests or measured performance. Provider APIs can change; check the linked documentation for your exact integration.