The agent is not done until every requirement has evidence.
AI agents often appear productive while missing a required step: a test was not run, a file was not updated, a URL was not verified, a branch was not checked, or a blocker was quietly assumed away. This checklist turns vague “did the agent do it?” reviews into a requirement-to-evidence audit.
- Restate the required outcome. What would be visibly true if the task were complete?
- List every required artifact. Files, URLs, reports, PDFs, CSVs, commits, cards, logs, or screenshots.
- Map requirements to evidence. Every requirement needs a file path, command output, screenshot, diff, API read, or public URL.
- Check for silent substitutions. Did the agent replace execution with a plan, or live verification with local-only checks?
- Check the negative constraints. No spend, no unauthorized accounts, no raw secrets, no unrelated brand references, and no customer-facing actions outside scope.
- Confirm verification was direct. Public URL fetched, download opened or returned 200, tests actually ran, data file exists, deployment state checked after deploy.
- Identify skipped-decision points. Was owner approval needed? Was a blocker converted into a safe fallback?
- Run a second-pass diff. Look for placeholder text, old brand names, broken relative paths, and empty sections.
- Record the final state. Done, partially done, blocked, or unsafe — with exact next action.
- Create one prevention rule. Add a prompt rule, test, checklist item, or cron instruction to prevent the same miss next time.
Default rule: no evidence means the step is not complete. “I verified it” is not evidence unless the agent names the command, URL, screenshot, API read, or exact file.
Requirement-to-evidence matrix
| Requirement | Expected evidence | Evidence found | Status | Fix needed |
|---|---|---|---|---|
| Deliverable created | File path + non-empty content | Pass / Fail | ||
| Public page updated | Live URL + visible change | Pass / Fail | ||
| Download works | HTTP 200 + correct filename | Pass / Fail | ||
| Brand separation preserved | Search result or grep output | Pass / Fail | ||
| Tests/checks run | Command + result | Pass / Fail | ||
| Blockers documented | Needs-from-owner note | Pass / Fail |
Prevention prompt
Before you finish, create a requirement-to-evidence table. For every requested deliverable or constraint, include the file path, command output, public URL, API read, screenshot, or exact blocker. If any requirement lacks evidence, keep working or mark it blocked with the smallest next ask. Do not claim verification from memory.
Red flags that mean the agent probably skipped something
- The final answer says “ready to deploy” but no deployment was attempted or verified.
- It reports a file path that does not exist.
- It says “verified” without naming the command, URL, screenshot, or API read.
- A generated page links to missing assets.
- A paid CTA points nowhere and no fallback contact path exists.
- A task references a prior brand, client, or workspace that should be separate.
- The agent did research only when the mission required building an asset.
Consultant monetization angle
Package this as an Agent Work QA Review: $150–$300 for a one-time audit, $500–$1,500/month for recurring agent-output QA, or $1,000–$3,000 to implement evidence gates, CI checks, deployment verification, and rollback documentation.