Verify AI Agent Work: Deliverables, Evidence, and Scope
Done may mean a file was created but never opened, or a post prepared but not published. Acceptance starts with defining completion and the evidence proving it.

Compare the agent’s report with the actual result
AI agents can plan, call tools, and perform multiple steps, but their final messages still need checking. Saying a file was created does not prove it opens; saying delivery succeeded does not prove the destination contains it. Accept work based on actual system changes or artifacts matching the request.
Anthropic’s discussion of agent evaluation distinguishes work logs from final outcomes. Use reports to identify what to check, then inspect the files or destination status. Verification does not require disclosure of every reasoning step.
This article proposes acceptance checks, not a ranking of agent trustworthiness. All examples are hypothetical, not records of actual service failures.
Agree on the criteria for completion before starting work.
Change “Make me a report” to “Create a Thai report that opens correctly, matches the attached figures, shows sources, and identifies unconfirmed items.” Then separate actions the agent can take independently from those requiring permission: drafting an email and sending it are separate steps.
| Type of work | Evidence to check | An insufficient status report |
|---|---|---|
| Documents | The file opens, content is complete, and figures and formatting match the brief | File created |
| Edit websites | The actual page renders, images and links work, and checks match the work version | The code looks correct |
| Prepare a post | Final text and images with draft or scheduled status | Ready to post |
| Authorized publication | Actual item in destination, time and content exactly as approved | Pressed send. |
Set criteria according to the consequences if wrong. Private drafts may be lightly examined. For work that will be released to the public or change important information, more detailed evidence is required. There is no need for every job to have the same number of inspection steps.
Check five aspects: real, correct, complete, usable, and within scope.
- Exists: The destination file or entry can actually be opened, rather than only showing a path name or confirmation screenshot
- Correct: Compare key information with originals, recalculate, and open cited sources
- Complete: Check each requirement, including exceptions and missing-data cases
- Usable: Try the steps your recipient would take, such as downloading, opening a link, or searching for an article on their mobile phone.
- Within scope: Check that you have not edited other work, sent information, or published more than ordered.
If reviewing with a table, give each row “Criterion / Evidence / Result / Revision.” Unverifiable items must remain pending. Combining unchecked items with passed ones misrepresents readiness.
Hypothetical example: An event budget report with a correct total but incorrect scope
Suppose the original data lists a venue cost of 3,000 baht and equipment of 2,000 baht, explicitly excluding travel. AI produces a report with a 5,000-baht total. The addition is correct and the file opens, but a heading saying “Total cost of the event” still goes beyond the evidence because travel costs are unavailable.
Rename the result “Total of documented items: 5,000 baht, excluding travel,” and leave a field for the task owner to confirm travel costs. The agent should not invent an amount to complete the report. This is a hypothetical review exercise, not an actual estimate.
If an agent says it checked amounts, ask which entries and sources. Another AI with the same missing information may add confidence without evidence.
Match test results to the delivered version
“Tested” should include test names, artifact versions, dates, actual results, and exclusions. Checks run before final edits apply to the earlier version; they do not automatically establish that the new file passes.
For websites, distinguish successful builds, Git pushes, and production deployment: one does not prove the next. For posts, distinguish drafting, scheduling, and publishing. For files, distinguish creation from opening and checking contents. Clear status names prevent drafts from being mistaken for completed external actions.
Automated checks can catch structural and link problems in writing but cannot replace factual reading. Model-based scoring needs criteria and human-reviewed samples; an AI score is not the final judgment of quality.
A handoff report template for further verification
Deliverables: final filenames and locations
Passed criteria: each with evidence
Tests actually run: Test name, job version, and results.
Unchecked or unconfirmed items: state the reasons and impact
External action: Still in draft / Scheduled / Published with actual evidence.
Work to be done: Who is responsible and next steps?
A handoff should provide work records and inspection evidence without requiring disclosure of internal reasoning. A plausible narrative alone is insufficient.
Evidence such as tool results and logs still requires checks of provenance and completeness. For important actions, the approach OWASP emphasizes permission enforcement and approval checks close to execution. A later handoff report does not replace prior authorization.
Return it for revision when evidence is missing without restarting the entire task
If issues arise, identify failed criteria and the evidence used, such as “The amount is correct, but the text calls it the total cost although travel is excluded.” This is more actionable than “Do it better.” Preserve passing parts and review areas affected by revisions according to risk.
Accept the work when the evidence can be inspected, meets the agreed criteria, and states the limitations. The agent’s final report should identify where each result is and what has been checked so the recipient can verify it.
Sources
- Anthropic: Demystifying evals for AI agents
- OWASP: AI Agent Security Cheat Sheet
- OWASP Cheat Sheets Series Team
Sources reviewed October 14, 2026, Thailand time. Thai examples and prompts are KrobClick proposals, not product tests. Features and settings vary by version/account. Check originals and outputs before real use.