How can I verify my AI agent actually did the work?
The short answer Ask for the receipts, then recompute them. For every step the agent claims, there should be a record minted when the step ran: the tool, the time, the saved result as canonical bytes, and a SHA-256 out
The short answer
Ask for the receipts, then recompute them. For every step the agent claims, there should be a record minted when the step ran: the tool, the time, the saved result as canonical bytes, and a SHA-256 output_hash over those bytes. Recompute the hash and compare. A match means the saved result was not changed after it was recorded. A step with no record is not "probably done"; in AER-1's closed outcome vocabulary it is absent, and you should treat the claim accordingly. Work run through the Zambo endpoint, https://zambo.dev/mcp , mints these receipts automatically, and checking any of them is free at https://zambo.dev/verify , no account, whether or not you paid for the work. Then run your own tests for the second, separate question: whether the work was any good.
Disclosure, once: I run ops for Zambo, the endpoint whose receipts this procedure checks.
Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.
The top answer today, and the half it covers
The leading result for this question is a solid DEV article about coding agents: put a test gate between "done" and merge. Run the suite in the agent's worktree, read the exit code, send the last fifty lines of failure output back for a retry, escalate after two failures. As far as it goes, it is right, and you should do it.
But watch what it verifies and what it leaves behind. The gate checks one step, the test run, in one place, your machine, at one moment, and its result is a boolean in a terminal that evaporates. It says nothing about the other steps the agent claims: the research it says it read, the file it says it checked, the analysis behind the change. It covers coding only; the agent that produced your market analysis or your client report has no test suite to gate. And a week later, when someone asks what actually ran, the gate has no record to show them. It answered "did the tests pass just now." The question on this page is bigger: did the work happen, all of it, and can anyone still check?
That is the half receipts cover. The two halves are complements, not rivals: the gate for correctness, the receipt for occurrence and integrity.
The checking procedure
1. Get a receipt id for every claimed step. Not a summary, not a log excerpt: an id per step, resolving to a public record. If the agent worked through the endpoint, 153 tools in total with 37 on the default surface, the ids exist by construction. If it worked elsewhere, ask what record it can point to. The answer sorts claims faster than any argument.
2. Recompute each one. Fetch the record, take its canonical bytes, hash them, compare with the output_hash. Equal: the saved result is the result that was recorded at run time. Not equal: something changed after the fact, and the conversation changes with it.
3. Name the outcomes honestly. AER-1 allows exactly five: verified, verification failed, rejected: unresolvable record, not checked, and absent. The last two do the quiet work. Not checked means a record exists and nobody has run the recompute yet, which includes you, right up until step 2. Absent means the step ran outside the receipted path, or not at all, and the record cannot tell those apart, which is precisely why work you care about belongs on the receipted path.
4. Then, and only then, judge quality. Run your tests, read the diff, check the citations. A verified receipt under bad work is verified bad work, and the receipt will tell you so itself if you read the saved result it protects.
Two receipts, recomputed in full
A data pull. On November 22, 2026, live_price returned BTC at $59,710.86 USD, up 3.07 percent over 24 hours, via Coinbase, fetched 2026-11-22T09:05:07Z. Receipt e25792a3-ce97-45a3-858b-78ee2fda8fbb: https://zambo.dev/run/e25792a3-ce97-45a3-858b-78ee2fda8fbb . The record carries 669 canonical bytes. SHA-256 over them: sha256:3e7d6beea897b25f6d09d02a7113c473a04fb54c5749e66169120458e75584e8, identical to the published output_hash. MATCH, outcome verified.
An analysis step. The same morning, doc_intel analyzed a real submitted document and minted receipt 34d47e29-1b53-40a9-84df-c9cac9841245 at 2026-11-22T09:05:56Z: https://zambo.dev/run/34d47e29-1b53-40a9-84df-c9cac9841245 . Its 3,580 canonical bytes recompute to sha256:0a8ceb6330503a1f181f4fc3b0089a663e7af5d74eb6edf805435abbaf5f2e37, matching its output_hash exactly. MATCH, outcome verified.
Two different tools, two different kinds of work, the same ninety-second check, neither of them requiring you to trust the agent, the vendor, or this article. That uniformity is the point: the format behind the receipts, AER-1, IETF Internet-Draft draft-zambo-aer1-26, ships a public conformance corpus, kit 1.5.7 with 143 test vectors and 146 checks, that independent implementations pass, so the checker you rely on does not have to be the maker's own software.
The honest limits
A receipt proves the saved result was not changed after the hash was recorded. It never proves the result was right. The price receipt above verifies identically in a world where the feed glitched that minute, and a receipted test run can be the wrong suite, on a stale branch, in a dirty environment. That is the test gate's territory, and this article is not arguing you out of it. Provenance is bounded too: a record states which tool the system says ran, a producer assertion, and the design leans on the recompute precisely because assertions are cheap. What verification gives you is a floor that did not exist before: occurrence and integrity as arithmetic, coverage gaps named absent instead of waved through, and a record that survives long enough for someone else to check your checking.
What it costs
Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. Verifying receipts costs the checker nothing at any tier, forever, which is what makes the procedure above usable on work you did not commission and agents you do not run. Paid tiers, stated once: Day Pass $0.99 unlimited for 24 hours, Zambo Pass $29/mo, $ZAMBO holder Pass $19/mo, Lifetime Pass $250. There is no annual tier.
Try it: Day Pass, $0.99 unlimited for 24 hours: https://zambo.dev/day-pass/ and add Zambo to your AI: https://zambo.dev/install/ (redirects to the add page, verified live today).
Sources
- Worked receipt one, live_price BTC $59,710.86 USD at 2026-11-22T09:05:07Z, 669 canonical bytes, output_hash sha256:3e7d6beea897b25f6d09d02a7113c473a04fb54c5749e66169120458e75584e8, recomputed MATCH, outcome verified: https://zambo.dev/run/e25792a3-ce97-45a3-858b-78ee2fda8fbb
- Worked receipt two, doc_intel at 2026-11-22T09:05:56Z, 3,580 canonical bytes, output_hash sha256:0a8ceb6330503a1f181f4fc3b0089a663e7af5d74eb6edf805435abbaf5f2e37, recomputed MATCH, outcome verified: https://zambo.dev/run/34d47e29-1b53-40a9-84df-c9cac9841245
- Public verification, no account required: https://zambo.dev/verify
- AER-1 conformance corpus, kit 1.5.7, 143 vectors, 146 checks: https://zambo.dev/aer1/corpus.json
- AER-1 draft text: https://www.ietf.org/archive/id/draft-zambo-aer1-26.txt
- Day Pass: https://zambo.dev/day-pass/
- Install: https://zambo.dev/install/
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.