Action State launches EvidenceBook to verify AI agent actions

Action State's entry into agent verification rests on a distinction its founder draws between seeing inside your own walls and trusting what happened beyond them. Most observability and tracing, Steven Mih said, concerns what was happening within a company's own systems. The new tools instead target transactions where one company cannot rely solely on another party's account of an agent's actions, a gap that widens as automated systems buy, sell and delegate work across corporate boundaries.

The design avoids a central authority holding everyone's data. Each participant keeps its own signed transaction histories, and comparing selected records can reveal agreement, conflicts or gaps. The company is explicit about the limits: preserving a record does not establish that its contents are accurate or that every action was captured, and Countersign does not certify that claimed outcomes occurred or that evidence is complete. Terms Packs, which define evaluation criteria and methods, do not certify legal compliance. Human review remains part of the workflow, with AI judges assessing records daily and employees periodically reviewing samples.

Early integrations include Mesh-LLM, an open-source distributed inference network, and Project Nanda, which used Agent Action Capsules in transactions between agents discovered through separate registries. EvidenceBook is available under the Apache 2.0 license, and its underlying specifications build on published Internet Engineering Task Force standards, though Action State's own specifications remain individual Internet-Drafts rather than an adopted standard. The company plans to seek neutral foundation governance for the specifications and implementations, and expects to earn revenue from independent verification services.