The News
Action State Group, a San Jose-based startup founded by former Ahana co-founder Steven Mih, announced EvidenceBook at AGNTCon + MCPCon North America. EvidenceBook is an open-source (Apache-2.0) evidence layer for AI agents that creates tamper-evident, cryptographically signed records of agent actions, enabling independent verification of what an agent did and whether it met defined expectations. The announcement also includes Terms Packs (structured requirement sets for evaluating agent work), the Countersign service (an independently operated second signature over evidence bundles), and integrations with open projects including Mesh-LLM and Project NANDA, all built on IETF standards SCITT (RFC 9943) and COSE Receipts (RFC 9942).
Analyst Take
The trust gap AI agents are about to expose at scale
There is a structural problem forming at the center of agentic AI adoption, and most enterprises have not named it yet. As AI agents gain the ability to choose suppliers, negotiate terms, execute purchases, and delegate work to other agents across organizational boundaries, the audit trail question becomes unavoidable: who is accountable when something goes wrong, and how do you prove what the agent actually did? EvidenceBook is an attempt to answer that question before it becomes a crisis.
According to ECI Research’s Google GovTech Survey, 45.6% of respondents say their organization permits AI agents only for approved use cases under defined policies, while a further 29.7% limit them to pilot projects or specific teams. That level of caution is rational. Without an independent evidence layer, organizations are essentially trusting the agent’s operator to self-report. That is not an audit; it is a courtesy. EvidenceBook’s bilateral record model, where a buyer’s agent and a provider’s agent each maintain their own tamper-evident capsule and can compare cryptographic references without centralizing control, is precisely the architecture that lets governance policies become technically enforceable rather than aspirational.
Why the architecture matters for developers
For engineering teams, the design choices here are worth examining closely. Action State builds on two published IETF standards, SCITT (RFC 9943) and COSE Receipts (RFC 9942), rather than inventing a proprietary ledger. That decision matters for portability. The Agent Action Capsule format at agentactioncapsule.org is an open record format with conformance vectors, meaning any vendor can produce compatible evidence. Countersign, the independently operated verification service, recomputes cryptographic and structural checks and signs exactly what it found, recording results in its own transparency log. It makes no claim about whether all actions were captured or that stated outcomes occurred, a deliberately narrow scope that avoids overpromising on what cryptographic verification can actually guarantee.
The integration with Mesh-LLM, an open-source distributed inference network, is the clearest production signal. In a peer-to-peer inference model where nodes collaborate without a central authority, each node can now keep a signed record and compare it with peers when both opt in. That is a genuine architectural advance for anyone building on decentralized AI infrastructure. The Project NANDA integration, where agents from two independent registries closed live transactions and each walked away with the same cryptographic bilateral record, is less a demo and more a proof that the open agent economy has a viable trust primitive.
What ITDMs need to weigh
For technology decision-makers, EvidenceBook’s immediate relevance is as infrastructure for governance, not as a product purchase. The open-source core is free. Countersign is free while early adopters come on board. The business question is whether your organization’s AI agent deployment roadmap is going to require cross-boundary accountability, and the honest answer for most enterprises is yes, eventually. ECI Research’s Google GovTech Survey found that 31.8% of respondents cite FedRAMP and compliance approval friction as the single largest blocker preventing widespread AI adoption in developer workflows. An open, standards-based evidence layer that produces auditable, independently verifiable records is exactly the kind of artifact that compliance and authorization processes will eventually demand for agentic workloads. Getting familiar with the specification now, before procurement and legal teams start writing requirements around it, is a low-cost hedge.
The vendor-neutral positioning is also strategically significant. Action State is a Silver Member of the Agentic AI Foundation and has stated plans to propose the specifications for neutral foundation governance. If that succeeds, EvidenceBook’s architecture could become the baseline expectation for interoperable agent evidence, the way OpenID Connect became the baseline for federated identity. Organizations that wait to engage until it is an RFP requirement will be playing catch-up on both technical integration and internal policy alignment.
Looking Ahead
The next twelve months will test whether open evidence infrastructure can attract the network density it needs to matter. A bilateral record system only provides meaningful verification when both parties adopt compatible implementations. Action State’s strategy of seeding the ecosystem through open-source integrations (Mesh-LLM, NANDA) and public Terms Packs (EU AI Act Obligations, an Outcomes Pack for contracted work) is sensible, but adoption will hinge on whether major platform operators, cloud providers, and SI partners build compatible evidence endpoints into their agent runtimes. Watch for announcements from hyperscalers and major agentic platform vendors about SCITT-compatible audit logging as a signal that the ecosystem is consolidating around this approach.
The longer arc here is regulatory. Governments and large regulated enterprises are not going to permit autonomous agent-to-agent commerce without an audit trail they can interrogate independently. Action State is positioning EvidenceBook as the foundational layer for that audit trail, and the open specification path through IETF is the right mechanism to drive broad adoption. If the AAIF governance proposal succeeds and a working group forms, the specification timeline accelerates significantly. Organizations building agentic platforms today should treat Agent Action Capsule compatibility as a near-term architectural requirement, not a future consideration. The window to influence the specification is open now.
Stay Ahead of Application Development Trends
Get weekly analyst insights, research notes, event coverage, and AppDevANGLE updates delivered directly to your inbox.
Subscribe for Weekly Insights
Join technology leaders, practitioners, and GTM teams following the trends shaping modern software delivery.
Looking for deeper research access?
Explore ECI Research reports, survey insights, and market analysis through the ECI Research Portal.
