Your agent finished the task and reported success. The API returned 200 OK. The dashboard is green. None of that is proof of agent work, and this article shows the check that is: a record of the call, saved by the system that ran it, that a stranger can recompute byte for byte without an account, a dashboard login, or trust in the agent.
The direct answer
Proof of agent work is a record with three properties. First, it was written at the moment the work ran, by the execution layer, not written afterward by the agent describing itself. Second, it holds the exact result bytes that came back, committed under a SHA-256 digest computed over the canonical form of the record. Third, anyone can fetch those bytes and recompute the digest. If the recomputed digest matches the saved one, the saved result was not changed after the call. If it differs by one character, the record fails loudly. That is the whole test, and it needs no one's permission to run.
A worked example, run today
This morning I ran a live price check through Zambo, the work layer built by Brennan Zambo, and then checked the receipt the way a stranger would. The call returned BTC at $61,621.38, a Coinbase snapshot fetched Fri, 20 Nov 2026 09:30:22 GMT. Its receipt is public: https://zambo.dev/run/0658f647-9a26-4363-a3b3-02173188c8dd
The recompute: fetch the receipt's canonical bytes (670 bytes for this one), hash them with SHA-256, compare against the digest the receipt carries as its output_hash. My recompute produced sha256:1c672475b2ee0d8aa940c28294c089bf1d84beeba9fb282a81a0065014857e42, which matches the saved output_hash exactly. The saved result is the result that call returned, unchanged. That check took one fetch and one hash, and it is the same check you can run on any of the 152,616 public receipts Zambo has issued (1,100 of them today, read from the live counters at 09:35 UTC on Nov 20, 2026).
Why read-backs and status levels are not proof
A popular approach grades an agent's self-report into levels: verified, verified-custom, acknowledged, attested-only, unverified. Read that ladder carefully. Every rung is the agent's own system reading its own store and reporting what it found. An attested-only rung is the agent attesting to itself. A verified rung means the agent's library checked the agent's records against the agent's rules. That is bookkeeping, and bookkeeping is useful, but it is not proof, because the party being checked controls the record being checked.
The recompute test has no such loop. The bytes are public. The digest is arithmetic. The checker needs nothing from the system that produced the record except the record itself, and the record either recomputes or it does not.
One honest boundary, stated the way the standard states it: a receipt proves the saved result was not changed after the call. It does not prove the answer was right. A price can be stale, a summary can be wrong, and the receipt will still recompute, because the receipt's job is integrity of the record, not correctness of the content. Anyone selling you a proof layer that claims both is selling you the second claim for free.
The standard, and what happened to it this week
The record format is AER-1, an open Internet-Draft (draft-zambo-aer1-26) on the IETF datatracker. Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard. The published conformance corpus holds 143 vectors and 146 checks. Laurent Lemonnier, an independent implementer who built his verifier from the draft text alone, passes 143 of 143.
On November 19, 2026, five separate systems reorganized around the standard in a single day. An independent implementer published draft v0.2 carrying four AER-1 answers as spec text, with the review credited. A governance toolkit maintainer confirmed the custody and reliance split holds in a system that actually runs it. A payments builder adopted the buyerAcceptance record and the resultDigest rule, deployed his v0.2 contract to Base Sepolia that night, and the recompute against the deployed contract came back clean. Two more implementers converged on the same key-recovery design in the same hour. A sixth followed the next morning, adopting the canonicalization profile into his own spec pull request. Proof of agent work stops being a vendor claim when other people's systems start bending around the record format.
Run the check yourself
Connect any MCP client to https://zambo.dev/mcp. 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, open its receipt link, fetch the canonical bytes, hash them, compare. A Day Pass is $0.99 for unlimited calls for 24 hours; the Zambo Pass is $29 a month, $19 for $ZAMBO holders, and Lifetime is $250. But the proof check itself costs nothing, and it works on receipts minted before you ever heard of Zambo. That is what makes it proof: it does not need you to buy anything, believe anything, or trust the agent. It needs a hash function.
Top comments (0)