The open standard for AI agent auditability

What is AAS-1?

AAS-1 is an open standard for producing audit-grade evidence about agent activity. It defines a record format, five record classes, and twelve standard assertions that allow auditors to evaluate agent actions against the same rigor applied to human-controlled processes.

The standard addresses what we call the Unauditable Agent Problem: autonomous agents are now executing transactions, making decisions and producing outputs at scale, yet the records they leave behind are platform-specific, unstructured and not designed for assurance. AAS-1 defines a portable evidentiary format, anchored to AIS-1 identities, that any auditor can evaluate.

At its core AAS-1 is a relationship between two parties — the agent emits a Class A record per action; the auditor emits a Class D opinion in response. Class B and Class C aggregate Class A for sampling; Class E wraps the engagement. The diagram shows the actor-output pair at the heart of the standard.

The AAS-1 Audit Manual has been prepared to assist practitioners in applying the standard — see the Audit Manual section below for details.

Five record classes. One spine.

A complete audit trail typically comprises many Class A records, organised by Class B or Class C aggregations, evaluated against a Class E engagement, and concluded with Class D determinations. Each class shares the core identity, evidence and signature primitives.

Twelve assertions. One evaluation pass.

AAS-1 records support seven assertions adapted from established audit frameworks (ISA 315/330, AICPA AU-C, ISAE 3000), plus five assertions specific to autonomous agents. An auditor — human or agent — evaluates every assertion in a single pass before issuing a Class D determination.

// Auditor verification flow — evaluating an AAS-1 Class A record

const record   = await aas1.fetch(eventId);
const identity = await ais1.resolve(record.agentRef);

const findings = await auditor.evaluate(record, {
  identity,
  assertions: AAS1.ALL_ASSERTIONS,
  policies:   record.policyRefs
});

// Returns per-assertion findings:
//   existence       'satisfied' | 'modified' | 'exception'
//   completeness    ...
//   accuracy        ...
//   identity        ...   (AIS-1 binding)
//   provenance      ...   (model, tools, prompt context)
//   reproducibility ...

if (findings.allSatisfied) {
  await auditor.issueClassD({ finding: 'unmodified', ... });
} else {
  await auditor.issueClassD({ finding: 'modified', exceptions: ... });
}

What is the Audit Manual?

The Audit Manual is the operational companion to the specification. Where the spec defines what AAS-1 records are, the manual sets out how an auditor receives them, evaluates them, and issues an opinion. It is the practitioner's playbook for conducting an AAS-1 Class D Determination.

A specification on its own would leave every auditor inventing their own procedures from the spec text. The manual closes that gap. It walks through pre-engagement scoping, receiving and validating the record population, the per-record verification flow, detailed procedures for each of the twelve assertions, sampling methodology for large populations, and the form of the Class D Determination itself. A worked example over the PayAgent USDC transfer shows the procedures applied end-to-end.

This makes AAS-1 immediately deployable rather than aspirational. An auditor can pick up the manual, follow the procedures, and conduct an engagement without further interpretation. The first AAS-1 engagement — over PayAgent's Q2 2026 records — is being conducted to this manual; subsequent engagements use it as the standard reference.

From draft to deployed standard.

Specification and supporting materials.

AAS-1 v0.1 is a draft for comment.

Feedback is invited from auditors, agent developers, blockchain engineers, legal and regulatory professionals, enterprise AI deployers, standards organisations and government bodies. The comment period closes 31 July 2026. A revised v0.2 will incorporate substantive feedback.