Part of the **Did My Agent Lie?* verification series — on proving what your AI agent actually did, not just what it said.*
To verify AI agent work, ignore the summary and check for independent evidence: confirm each claimed action has a verifiable receipt from the execution layer, spot-check outputs against real systems — the file exists, the row was written, the email sent — and rerun anything cheap. If the proof is the agent saying "done," you've delegated trust to the narrator.
The verification ladder
Not all evidence is equal. Think of verification as a ladder — each rung is harder for an agent to fake than the last:
| Evidence level | What it is | How hard to fake |
|---|---|---|
| Chat claim | "I finished the task." | Trivial — it's just text. |
| Transcript | The agent's full action log, as it reports it. | Easy — still the narrator's own story. |
| Receipt | Execution-layer record: tool, inputs, outputs, timestamp. | Hard — generated outside the agent's storytelling. |
| Reproduced state | You open the file, see the record, get the email. | Practically impossible — it's reality. |
Your job is simple: climb as high as the stakes demand. Low-stakes draft? The transcript may be fine. Money moved? You want verifiable receipts and reproduced state.
This is the forensic version of verification — the evidence ladder, the reproducibility tests, the parts that catch an agent in the act. If you want the plain-language version first, start with the companion piece in this series, How Do I Know If My AI Agent Actually Did the Task, then come back here for the full workup.
The checklist
1. Demand the execution log, not the summary
Ask: "Show me every action you took, with timestamps." If the agent can only produce a summary and no log, that's your first red flag — you're being asked to trust the narrator.
2. Match every claim to a verifiable receipt
Go through the agent's summary line by line. Each "I did X" needs a corresponding executed action in the log. Claims without matching actions are unproven until proven. A verifiable receipt — a tamper-evident record from the execution layer naming the tool, inputs, outputs, and time — is the strongest match you can get.
3. Spot-check the critical path in the real world
You don't have to redo everything. Pick the one or two claims the whole task hinges on and verify them yourself: open the file, query the record, check the sent folder. One real-world confirmation is worth a hundred confident sentences.
4. Re-run anything cheap
If a step is cheap to repeat — a search, a calculation, a read — re-run it. Genuine work reproduces; fabricated work doesn't. This is the fastest way to catch a hallucinated completion.
5. Keep the trail
Save the logs and verifiable receipts for anything consequential. Verification isn't just about this task — it's about the next one, when you'll want to compare what "done" looked like then versus now. Agents without an audit trail are agents without accountability.
What good verification feels like
It feels boring. That's the point. You run the checklist, the evidence lines up, you move on with your day. The drama only happens when something doesn't line up — and that's exactly when you want the drama: early, while you can still fix it. An agent that gets verified every time gets better at producing verifiable work, because it learns (in the practical sense) that "done" is a claim that will be checked.
Frequently asked questions
What is the simplest way to verify AI agent work?
Ask for the execution log — the timestamped record of every tool call with inputs and outputs — and match each claim in the agent's summary against it. Then spot-check the single most important claim in the real world yourself. If summary and evidence agree, the work is verified.
Should I trust my AI agent's summary of what it did?
Treat it as a status report, not proof. Summaries are generated text; they can drift, skip steps, or narrate the plan instead of the execution. Verification means finding evidence that does not come from the agent's own words — verifiable receipts, timestamps, and changed state.
What is a verification receipt for AI agent work?
A verifiable receipt is a tamper-evident record created by the execution layer showing that a specific action really ran: the tool, its inputs, its outputs, and the time. Receipts prove execution integrity — that the work really ran — never correctness. For the full breakdown, read the companion piece in this series, What Is an AI Execution Receipt.
How do I verify agent work on a task I can't redo myself?
You don't need to redo the whole task. Verify the weakest link: pick the one or two claims the task's success depends on and check those against independent evidence. Require verifiable receipts for the rest. Partial verification of the critical path beats full trust of everything.
How often should I verify my AI agent's work?
Verify consequential actions every time — anything that spends money, changes data, sends messages, or commits you to something. For low-stakes work, random spot-checks keep the agent honest. The rule of thumb: the costlier a mistake, the stronger the evidence you require before accepting "done."
Verify it yourself
Want this checklist run automatically? Paste your agent's transcript into Did My Agent Lie?, a free check from Zambo — it matches every claim against verifiable receipts.
Top comments (0)