DEV Community

Cover image for What is AER-1?
rambo
rambo

Posted on

What is AER-1?

AER-1 is the open standard for agent execution receipts. Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard.

The problem: "done" is a claim, not a record

When an AI agent finishes a job, the evidence is usually a log the agent wrote itself, or a chat message saying it worked. Neither can be checked by anyone else. If the agent says it fetched a price, sent the file, or deleted the record, you are trusting the narrator.

AER-1 replaces the claim with a record. For every tool call, the exact output bytes are saved, and a SHA-256 hash of those exact bytes is published with the record. Anyone can recompute the hash and check that the saved result was not changed after the fact. A receipt does not prove the answer was right. It proves the record you are looking at is the record the call actually produced.

How an AER-1 receipt works

  1. A tool runs. Its name and version are bound into the record, for example live_price v4.1.4.
  2. The exact output bytes, the canonical bytes, are saved together with the inputs.
  3. The commitment is output_hash: SHA-256 over the canonical bytes. Change one byte and the hash breaks.
  4. Receipts chain per job, so a whole session of calls can be wrapped into one handover record a client or a second agent can pick up.
  5. The receipt id is a label outside the committed payload (Sections 6 and 13 of the draft). The commitment covers the work, not the label.

The outcome vocabulary is fixed by the spec, not free text. A call either executed or it did not, and the record says which. A refused or rejected call still mints a receipt with executed set to false, so "the call never ran" is itself a checkable fact instead of an excuse. Verification outcomes are fixed too: verified, verification failed, rejected: unresolvable record, not checked, absent. Two verifiers reading the same record reach the same verdict words.

The spec

AER-1 is an IETF Internet-Draft, draft-zambo-aer1-26, closed at revision -26. Read it here: https://datatracker.ietf.org/doc/draft-zambo-aer1/26/

The conformance corpus: do not trust, run it

A standard is only real if a stranger can test against it. The AER-1 conformance corpus is 143 test vectors and 146 checks, published with its digest: sha256:60bf73f325793e16843370cfa24903bb0d017525cb297e42bebc35857e2db0d2. Independent builders have written their own verifiers against it in Node.js, Java, Perl, and Python. Five independent validators are listed on the public registry, one of them scored 143/143 against the full corpus, and the newest independent interop implementation shipped today, November 15.

Run the challenge yourself:

Run the corpus against your own verifier and post your score. A clean run is 146/146 checks.

A worked verification, run tonight

At 23:52 UTC tonight, November 15, I called live_price for BTC through Zambo. The call returned $65,235.37 and minted this receipt: https://zambo.dev/run/124128be-4901-4b53-b613-3ff99235d0a5

Here is the entire verification:

import base64, hashlib, requests

receipt = requests.get("https://zambo.dev/api/receipts/124128be-4901-4b53-b613-3ff99235d0a5").json()
raw = base64.b64decode(receipt["canonical_bytes"])
print(len(raw), "canonical bytes")
print("sha256:" + hashlib.sha256(raw).hexdigest())
print("matches output_hash:", "sha256:" + hashlib.sha256(raw).hexdigest() == receipt["output_hash"])
Enter fullscreen mode Exit fullscreen mode

Output:

671 canonical bytes
sha256:d1fe1a4dbfc9ae6ea02e7b5945c59b38b7e84e9b8089372d6697ceae1b7f0778
matches output_hash: True
Enter fullscreen mode Exit fullscreen mode

No account, no trusting the agent, no asking anyone. The bytes either hash to the commitment or they do not.

AER-1 in production

As of tonight, November 15: 140,361 public receipts (2,249 of them minted today), 120,572 tool calls, 270,700 package downloads, and 430 workflow seals. Every one of those receipts can be recomputed the same way as the one above. That is the difference between a standard with a PDF and a standard with traffic.

Try it

Top comments (0)