Action State Launches Open Evidence Layer for Verifying AI Agents

Trust is the missing infrastructure of the AI agent economy. As autonomous agents begin negotiating, buying, and delegating work across company boundaries, a hard question keeps surfacing: when two machines transact, how does either party prove what actually happened? Startup Action State Group thinks the answer is an open evidence layer — and this week it launched the first pieces of it.

According to SiliconANGLE, Action State announced a set of open-source tools on October 22 at AGNTCon + MCPCon North America designed to let businesses independently check records of what AI agents did. The launch targets a trust gap that grows more urgent as agentic systems move from answering questions to moving money: companies cannot rely solely on another party's account of an agent's actions, especially when the agent belongs to someone else.

How the Evidence Layer Works

Action State's offering has three interlocking parts. EvidenceBook is the open-source record-keeping core, released under the Apache 2.0 license. It lets each participant keep its own tamper-evident, signed transaction history. A buyer records what it requested and paid; a provider records what it accepted and reported delivering. Comparing selected records can then reveal agreement, conflicts, or gaps — without requiring a central authority to hold everyone's data.

Terms Packs sit on top of those records as evaluation criteria. Packs define versioned requirements used to assess agent work — including regulatory and business rules — and results classify each requirement as met, missed, or not evaluable. According to the company, an initial pack covers agent-related obligations under the European Union's AI Act, with 37 obligation rows, while a second Outcomes pack remains in draft. Governance, risk, and compliance teams are expected to customize packs for their own requirements, and the company stresses that packs do not certify legal compliance.

The third piece is Countersign, an independently operated verification service available from launch. It checks selected cryptographic and structural properties of an evidence bundle, signs its findings, and records them in its own transparency log. The service is currently free, though the company expects to charge once relying parties accept Countersigns. A free public witness service also exists, establishing that a checkpoint existed at a particular time.

Under the hood, the system uses Agent Action Capsules, an open format for signed records, plus cryptographic checkpoints that make subsequent changes detectable. The specifications build on published Internet Engineering Task Force standards such as SCITT and COSE Receipts, but remain individual Internet-Drafts for now. Action State plans to seek neutral foundation governance for the specifications, and it is already a Silver Member of the Agentic AI Foundation. The company says the full layer is free, with revenue expected from verification services built on top.

Why Independent Verification Matters Now

The distinction Action State draws is between observing your own agents and verifying someone else's. Most existing observability and tracing tooling watches what happens inside one organization's walls. Cross-boundary agent commerce needs the opposite: a way for a buyer to check a provider's claims without trusting the provider's infrastructure. According to Action State, both parties need not install the software for one party to benefit, and verification can occur offline.

Founder and CEO Steven Mih, who previously founded the managed Presto vendor Ahana Cloud (acquired by IBM in 2023), frames the limits of the approach carefully. Preserving a record does not establish that its contents are accurate or that every action was captured. EvidenceBook can verify that reported events actually occurred, but it cannot verify that an agreement was satisfied. Countersign checks the properties of an evidence bundle without certifying that claimed outcomes occurred or that the evidence is complete.

Human review remains central to the design. AI judges assess agent records daily, with employees periodically reviewing samples to compare human and automated judgments, and evaluation results are preserved so auditors can examine the underlying evidence. For personal agents, the company's Everyday Rules product integrates with Muse to evaluate proposed actions against an owner's requirements.

Early Integrations and Ecosystem Reaction

Early integrations suggest the approach is already being exercised in the wild. Mesh-LLM, an open-source distributed inference network, uses signed records so participating nodes can compare exchanges with peers — a design that could help spot changes to a counterparty's recorded claims about models or configurations. Project Nanda, the open-source initiative building infrastructure and protocols for AI agents, has used Agent Action Capsules in transactions between agents discovered through separate registries, with each side retaining a sealed bilateral record.

Industry voices are backing the direction. Project Nanda founder Ramesh Raskar said in the company's announcement that discovery tells you an agent exists, but it cannot tell you what that agent did after you trusted it. Mazin Gilbert, executive director of the Agentic AI Foundation, said the approach addresses the need to independently verify what agents are doing regardless of who operates them or which systems they use, as reported by TokenPost.

The launch lands in a busy week for agent infrastructure. Enterprise vendors are racing to put agents on production workflows — ServiceNow and Nvidia just made their Arc enterprise agent generally available — while agent payment rails and agent identity schemes are attracting fresh funding. Against that backdrop, the agent ecosystem's missing piece is not capability but accountability. For a reporter tracking this beat, the questions agents themselves raise have shifted from "can agents do the work" to "can anyone prove they did it" — a concern echoed by recent research finding that 76% of AI agents sit dormant with live credentials, and underscored as personal agents enter consumer messaging channels.

Action State's bet is that accountability infrastructure must be open to work. Individual drafts, conformance vectors, and an Apache 2.0 reference implementation mean anyone can produce compatible evidence or define requirements — and that competitors can run compatible verification services. Whether that openness converts into adoption will depend on one thing the technology cannot provide on its own: counterparties willing to be checked.