Action State open-sources EvidenceBook to make AI agent work independently checkable

Action State Group on October 22 launched a set of open-source tools meant to let businesses independently check what AI agents actually did, as automated systems begin buying, selling and delegating work across company boundaries.

The package has three parts: EvidenceBook, which preserves tamper-evident records; Terms Packs, which encode how agent behavior should be evaluated; and Countersign, an independent verification service.

The problem is concrete. When an agent acts across an organizational boundary, one party cannot rely solely on the other's account of what happened. EvidenceBook gives each side its own signed transaction history — a buyer records what it requested and paid, a provider records what it accepted and reported delivering.

Comparing selected records can reveal agreement, conflict or gaps without a central authority holding everyone's data. Participants keep control of their information and can disclose more when evaluating outcomes; one side can benefit even if the other has not installed the software, and compatible implementations can exchange evidence directly.

Technically the system uses Agent Action Capsules, an open format for signed records, plus cryptographic checkpoints that make later changes detectable and a free public witness that establishes a checkpoint existed at a given time. Verification can run offline and costs nothing.

Founder and CEO Steven Mih, previously the founder of Ahana Cloud, was explicit about the limits: preserving a record does not establish that its contents are accurate or that every action was captured. EvidenceBook verifies that reported events occurred, not that an agreement was satisfied, and Countersign checks cryptographic and structural properties without certifying outcomes. Terms Packs go further on requirements — an initial pack covers agent-related obligations under the EU AI Act — but do not certify legal compliance.

Human review stays in the loop. Mih described AI judges assessing agent records daily while employees sample and compare judgments, with evaluation results preserved so auditors can examine the underlying evidence. EvidenceBook ships under Apache 2.0, early integrations include Mesh-LLM and Project Nanda, and the specifications remain IETF Internet-Drafts rather than RFCs; Action State says it will seek neutral foundation governance and plans to charge for Countersign once relying parties accept it.