You can add verifiable execution to an AI agent in an afternoon with a vendor SDK. What you get at the end of the afternoon is a record format one company controls, verified by that company's tools, readable while that company keeps the lights on. Verifiable AI agent execution that outlives a vendor is a different thing: an open standard, published as an IETF Internet-Draft, with a public conformance corpus that independent implementers pass without asking anyone's permission. That standard exists. It is AER-1, and this article shows its receipts, including one minted and recomputed this morning.
What the standard specifies
AER-1 (draft-zambo-aer1, closed at revision -26) defines the record one AI agent tool call leaves behind: the tool, the caller scope, the timestamp, the canonical bytes of the result, and a SHA-256 digest over the canonical receipt bytes. Verification is one recompute: fetch the bytes, hash them, compare with the saved digest. A match proves the saved result was not changed after the call. The standard is explicit about the limit: the check proves the record is intact, never that the result was right.
The conformance corpus is public: 143 vectors, 146 checks. The reference implementation is Zambo, built by Brennan Zambo. Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard. The corpus is the part vendor SDKs skip: anyone can build a verifier from the draft text alone and be graded against the same vectors.
Independent proof, named, dated, linked
The standing corpus fact first: Laurent Lemonnier built a verifier from the draft text alone, and his 143 of 143 against the full corpus is the only documented full-corpus pass. No kit internals, no vendor help.
Then, on Nov 19, 2026, five separate systems reorganized around the standard in a single day:
- Nilo published draft v0.2 of his own specification carrying four AER-1 answers as spec text, the review credited, his test vectors verified from the published bytes: https://njump.me/effbde77c76a6966552fb2e9127f6a425b5ad981ba6ead079bc566b7ef4e66ae
- John B Joseph confirmed on the agent-governance-toolkit discussion that the custody and reliance split holds in a system that actually runs it: https://github.com/microsoft/agent-governance-toolkit/discussions/4564
- Cullen Meyers wrote the same custody and reliance distinction into BoundaryAttest's own documentation through a merged pull request on that thread.
- kemppel adopted the buyerAcceptance attestation and the resultDigest rule on the x402 payments thread, then deployed his v0.2 contracts to Base Sepolia that night (contract 0xC0A315b5C4745DecB7db9Ed019C026971500e7a0, Sourcify exact match). From this desk, the recompute against the deployed contract matched across the board: https://github.com/x402-foundation/x402/issues/4324#issuecomment-7026482597
- verify402 and Payload converged on the same key-recovery design in the same hour on the x402 thread, and Payload reproduced the double-charge guard live, the payer balance moving exactly once ($0.05): https://github.com/x402-foundation/x402/issues/4293#issuecomment-7022131754
And on Nov 20, a sixth: Shikhyy updated his specification pull request to adopt the canonicalization profile and a new named outcome for an origin that cannot be canonicalized: https://github.com/x402-foundation/x402/issues/4315#issuecomment-7032965433
Standards are not adopted in press releases. They are adopted when other people's systems reorganize around them, in public, with links. Five in one day, a sixth the next morning, every one of them linked above.
This morning's receipt, recomputed
A live price call at 09:30 UTC today returned BTC at $61,621.38. Receipt: https://zambo.dev/run/0658f647-9a26-4363-a3b3-02173188c8dd. Canonical bytes: 670. Recomputed SHA-256: 1c672475b2ee0d8aa940c28294c089bf1d84beeba9fb282a81a0065014857e42, matching the saved output_hash. Zambo's public stream holds 152,616 such receipts, 1,100 of them minted today (counters read 09:35 UTC, Nov 20, 2026), every one checkable the same way at https://zambo.dev/verify/ with no account.
The SDK question, answered plainly
If a tutorial shows you verifiable execution through one vendor's SDK, ask three questions. Can a competitor's verifier check the record without the vendor's library? Can the record be checked after the vendor pivots, sunsets, or paywalls the endpoint? Does a public corpus grade the implementation, or does the vendor grade itself? An SDK that fails those questions gives you verifiable execution inside one company's garden. AER-1 was written so the record survives the garden.
Try the standard, not the pitch
Connect 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. 153 tools, 37 on the default surface. Make a call, open the receipt, recompute the digest with your own hash function, and then read the draft at the IETF datatracker and check the corpus yourself. The Day Pass is $0.99 for unlimited calls for 24 hours when a job outgrows the free day; the standard itself costs nothing to implement, which is the entire point of a standard.
Top comments (0)