Action State launches open evidence layer for AI agents

Action State unveiled EvidenceBook at AGNTCon + MCPCon North America to let AI agents keep tamper-evident records of what they did and what they agreed to do. The open-source system is designed to help buyers, providers and auditors verify agent work without relying on one operator’s account.

Why it matters: - AI agents are moving beyond single-company workflows and into transactions with other agents, vendors and customers. - Action State is aiming to make those interactions independently checkable, so organizations can verify what happened, what was agreed and whether work matched expectations. - The company is positioning evidence, not trust in an operator, as the basis for cross-company agent work.

What happened: - Action State announced EvidenceBook at AGNTCon + MCPCon North America on October 22, 2026. - EvidenceBook is an open-source evidence layer for AI agents released under Apache-2.0. - Action State also introduced Terms Packs and Action State Countersign, which is available today. - The company said the layer is free, with open specification drafts underneath, so others can produce compatible evidence, define requirements and verify records.

The details: - EvidenceBook lets an agent keep a tamper-evident record of what it did and what it agreed to. - The records are checkable by anyone without trusting the operator. - EvidenceBook uses Agent Action Capsules, the open record format at agentactioncapsule.org, and commits signed records to a tamper-evident history that independent software can verify. - The software ships as open-source libraries with A2A and MCP bindings, so it can run inside an existing agent harness or runtime. - A free public witness can show that a checkpoint existed by a particular time. - Participants can compare cryptographic references while keeping control of their data and disclose more only when an outcome must be evaluated. - Verification shows what the records say and that they have not changed. It does not prove the described events occurred. - Terms Packs are versioned lists of requirements that run over the records in the book. - Recomputed rows are facts or rules over the record, so anyone running them gets the same answer. - Judged rows rely on a judge, which can be an AI model pinned to a version or a named person. - The judging steps ship as skills, so the reasoning path is itself readable. - Results show whether each requirement was met, not met or could not be evaluated, along with the evidence behind each finding. - The evaluation itself is sealed as a signed record. - The same libraries apply a Terms Pack, and the host decides whether an action proceeds. - Public packs available today include an EU AI Act Obligations Pack and a draft Outcomes Pack for contracted work. - Everyday Rules for personal AI agents ships with the Muse integration and evaluates an action before it happens: allowed, not allowed, needs your approval, or not fully checked. - Packs do not certify compliance or score trust, and anyone can publish one. - Countersign runs checks as a party independent of both sides. - Countersign recomputes specified cryptographic and structural checks over an evidence bundle, signs exactly what it checked and found, and records the result in its own transparency log. - The finding can be carried to an auditor, insurer or court that was not present for the original transaction. - Countersign does not certify that all actions were captured or that claimed outcomes occurred. - Countersign is free while the first businesses that rely on it come on board. - The service is available at countersign.actionstate.ai.

Between the lines: - Action State is trying to separate proof from platform control. - By making the evidence layer open and portable, the company is inviting competitors and partners to build compatible services on the same standards. - The approach also draws a clear line between record verification and factual truth, which matters in disputes and audits. - The product set appears designed for situations where two sides need a shared record but do not fully trust each other. - The company is also leaning on open standards and IETF building blocks to increase the odds of broader adoption.

What happened in early deployments: - Mesh-LLM, an open-source distributed inference network, is using Action State’s portable evidence tools on participating nodes. - Nodes verify their own records and can exchange evidence with peers when both sides opt in. - Project NANDA agents discovered through the NANDA Index completed transactions with agents from an independent registry in a live deal room built on Agent Action Capsules. - Each negotiation produced a sealed bilateral record that both agents hold, and the integration merged into the NANDAtown codebase in July 2026. - Mih said at the conference session that Muse agents will buy and sell on the NANDA Index, with each action checked against the rules its owner set.

What’s next: - Action State plans to propose its specifications for neutral foundation governance. - The company is a Silver Member of the Agentic AI Foundation, but membership does not mean the proposed projects are accepted. - Mih is scheduled to present “Who, May, Did” on October 22 at 4:40 p.m. PDT in Room 230C. - Action State says availability includes open source and specifications at agentactioncapsule.org, on Action State GitHub repos and products at actionstate.ai. - Feature coverage varies by integration.

The bottom line: - Action State is betting that AI agent commerce will need portable evidence as much as it needs capability, and it wants that layer to be open, verifiable and vendor-neutral.