How do I check an AI agent's tool execution? Recompute the receipt, not the log
Your agent says the tool ran. It shows you a log line it wrote itself and a 200 status. Neither of those is a check. A log line is a claim from the same agent you are trying to verify, and a 200 only says the transport a
Your agent says the tool ran. It shows you a log line it wrote itself and a 200 status. Neither of those is a check. A log line is a claim from the same agent you are trying to verify, and a 200 only says the transport answered. The check that actually works is different: recompute the record of the call from the bytes that were saved at the time it ran, and compare what you get against what the receipt says.
That is what an agent execution receipt is for. Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard. Every tool call through Zambo leaves one, and the receipt is checkable by anyone who can read it, no account and no trust in the agent required.
The check, in four steps
1. Find the receipt. Every call returns one. It looks like https://zambo.dev/run/594cd425-6617-497a-aa33-a3766e165239. If the agent cannot produce a receipt for the call it claims it made, the honest outcome is "not checked". That is a recorded outcome, not a pass.
2. Fetch the receipt record. The record carries the tool name, the step status, the timestamps, the output hash, and the canonical bytes of what was saved. For the call below, the record shows tool live_price, status success, started 19:12:18.729Z and completed 19:12:18.902Z on November 20, 2026.
3. Recompute the hash yourself. Take the canonical bytes, run SHA-256 over them, and compare the result with the output_hash field. This is the whole check. If the saved result was changed after the call, the hash does not match.
4. Read the verdict. The receipt for the call below reports verification_status: commitment_matches, which means the stored bytes hash to the stored output hash. The other possible outcomes are equally plain: verified, verification failed, "rejected: unresolvable record", "not checked", and "absent". Never accept a sixth outcome invented on the spot.
Worked example, run today
At 19:12:17 GMT on November 20, 2026, I called live_price for BTC through the live Zambo endpoint at https://zambo.dev/mcp. The call returned:
BTC, $58,200.13 USD, down 6.66 percent in 24 hours, market cap $1,173.18B, snapshot via Coinbase.
The receipt: https://zambo.dev/run/594cd425-6617-497a-aa33-a3766e165239
Recomputing it from the saved record: 662 canonical bytes, output_hash sha256:8158542eec77e48f4544c12031d3447172f07c006bc68a88305c167580d9fe24. SHA-256 over those exact bytes gives the same value, character for character. MATCH.
Then the negative test, which is what makes the positive one mean anything: flip a single byte in the canonical bytes and recompute. The hash no longer equals output_hash. A changed record fails the check, and there is no way to edit the bytes and keep the match.
What this catches, and what it does not
It catches the common case: an agent that says it ran a tool when it did not, or that shows you a different result than the one the tool actually returned. No receipt, or a receipt whose bytes do not hash to its output hash, ends the question in your favor.
It does not prove the tool's answer was true. The receipt proves the saved result was not changed after the call; it does not prove the price was right. That distinction is the point. A check that claims more than it can do is worse than no check at all.
Try it on your own call
Every tool free once a day, plus 3 continue calls. One free wrap-up handover per day, and pickup is always free. Make one call, take the receipt link it returns, fetch the record, and recompute the hash. If you are past the free calls, the $0.99 Day Pass opens unlimited calls for 24 hours, and the same check applies to every call you make under it.
Today the counter sits at 154,221 receipts issued all time, 2,319 of them today. Every one of them is checkable the same way this one was.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.