Direct answer
What should I look for when choosing a tool to verify AI agent work?
Choose by who can recheck the record. The right tool writes the record while the work runs, lets a third party recompute the check without trusting the tool or the agent, shows failed steps instead of hiding them, exports everything, and gives your client a link they can open with no account.
By Brennan Zambo, creator of Zambo. Updated October 17, 2026.
The six tests
| Test | How to run it | What passing looks like |
|---|---|---|
| 1. Written at execution time | Ask when the record is created: during the call, or written up afterwards? | The record is saved by the connection as each step runs. A write-up produced later is a report, and reports can be edited. |
| 2. Recheckable by a third party | Hand the record to someone with no account and ask them to verify it. | They can recompute the check in a browser and reach the same verdict you did, without your login and without trusting you. |
| 3. Independent of the agent | Ask who writes the record: the agent being checked, or something else? | A separate connection writes it. A record the agent writes about itself fails this test by construction. |
| 4. Failures stay visible | Run a job with a step that fails and see what the record shows. | The failed step is in the sequence, marked as failed, next to the steps that succeeded. A record of only successes is a highlight reel. |
| 5. Honest limits, stated | Ask the tool what its record does not prove. | A straight answer: the record proves the saved result was not changed since it was saved; it does not prove the result was correct. |
| 6. Export and exit | Export a finished job and open the export somewhere else. | The full record leaves with you in a readable form, so your proof survives cancelling the tool. |
How Zambo does on its own tests
One: every call through the Zambo connection writes a verifiable AER-1 receipt as it runs, per step. Two: the receipt link opens a public check anyone can recompute at zambo.dev/verify, and the format is published as IETF Internet-Draft draft-zambo-aer1-23, so independent implementations run the same verification. Three: the connection writes the record, not the agent. Four: failed and refused steps are receipts too, kept in sequence. Five: the limit is stated on every page of this hub, including this one. Six: handovers export as PDF and JSON on the paid tiers, and receipts stay readable by link. Pricing is part of the test too: free to start, $0.99 for a 24-hour Day Pass, $29 a month for the Zambo Pass, and the specifics are on the cost worksheet page.
The honest part
No verification tool proves an AI's answer was right, and a tool that will not say so plainly is telling you something about its record as well. What a good tool proves is narrower and more useful: the steps that ran, the results that were saved, and that neither was quietly changed afterwards.
The three kinds of tools that verify AI agent work
Tools that verify AI agent work fall into three kinds. Evals test whether an agent's answers are good, usually on sampled tasks before or after a run. Traces and observability tools record what happened inside a run so your own team can debug it, and you read them on the tool's terms. Execution receipts fix each step's saved result under a commitment a third party can recompute, so someone who does not trust you or your stack can still check the work. The kinds answer different questions: evals ask was it good, traces ask what happened, receipts ask can anyone prove it. Client-facing proof needs the third kind, and Zambo is built on it: every call leaves a verifiable receipt, and the receipts roll up into a handover link the client checks directly.
Frequently asked questions
The three kinds of tools that verify AI agent work
Tools that verify AI agent work fall into three kinds. Evals test whether an agent's answers are good, usually on sampled tasks before or after a run. Traces and observability tools record what happened inside a run so your own team can debug it, and you read them on the tool's terms. Execution receipts fix each step's saved result under a commitment a third party can recompute, so someone who does not trust you or your stack can still check the work. The kinds answer different questions: evals ask was it good, traces ask what happened, receipts ask can anyone prove it. Client-facing proof needs the third kind, and Zambo is built on it: every call leaves a verifiable receipt, and the receipts roll up into a handover link the client checks directly.
What should I look for when choosing a tool to verify AI agent work?
Six things. The record is written while the work runs, not reconstructed afterwards. A third party can recheck it without an account and without trusting you. The record is written by a connection independent of the agent being checked. Failed steps stay visible in the sequence. The tool states plainly what its record does not prove. And the full record exports with you if you leave. Any tool that fails the first two tests is selling you a display, not a check.
What is the difference between a log and a verifiable receipt?
A log is a record a system keeps about itself; you read it on the system's terms and trust that it was not edited. A verifiable receipt binds each step's saved result with a commitment a checker can recompute later, so an edit after the fact breaks the check in the open. Logs are for watching work. Receipts are for settling questions about work after it is done.
Should the tool say what it does not check?
Yes, and treat silence on this point as a failed test. A verifiable receipt proves the saved result was not changed since it was saved. It does not prove the output was correct, and it says nothing about work that ran outside the recorded connection. A tool that names its limits lets you rely on it for exactly the questions it can settle.
Can my client check the record without an account?
On Zambo, yes. The receipt and handover links open in a browser with no account, no install, and no AI tool on the client's side, and the client can run the verification and record an Accept or Flag decision on the same page. Recipient friction is where verification usually dies, so test it with a real second person before you commit to a tool.
What should I test before paying anything?
Run one real job end to end on the free tier: do the work, open the record, hand the link to someone else, and watch them check it without your help. On Zambo the free demo takes a minute and the free tier covers up to 20 free calls per metered tool per day. If the free run does not produce a record a stranger can check, a paid plan will not fix that.
Your AI said "done." For $0.99, prove it: https://zambo.dev/day-pass/
Buying, not browsing? The main site turns this into a buyer checklist: what to look for when choosing a tool to verify AI agent work.
What tools exist to verify AI agent work?
Three kinds. Evals score whether an agent's answers are good. Tracing and observability tools record a run so your own team can debug it. Execution receipts fix each step's saved result under a hash commitment anyone can recompute, which is the kind a client or auditor who does not trust your stack can check. For proof you can hand to someone else, you want receipts: on Zambo every call leaves one, and a $0.99 Day Pass covers a day of receipted work.