Startup Action State launches tools for verifying interactions between AI agents

Startup Action State launches tools for verifying interactions between AI agents 

Action State Group Inc. is launching a set of open-source tools designed to let businesses independently check records of what artificial intelligence agents did, addressing a trust problem as automated systems begin buying, selling and delegating work across corporate boundaries.

The offering combines EvidenceBook, which preserves tamper-evident records, Terms Packs for evaluating agent behavior and Countersign, an independent verification service. The approach targets transactions in which a company cannot rely solely on another party’s account of an agent’s actions.

The fundamental difference between the start-up’s approach and that of observability vendors “is that most of the observability and tracing has to do with what was happening within your own walls,” said Steven Mih, Action State’s founder and chief executive. Mih was previously founder of Ahana Cloud Inc., the developer of a managed version of the Presto open-source distributed query engine. IBM Corp. acquired Ahana in 2023.

EvidenceBook gives participants their own signed transaction histories. A buyer can record what it requested and paid, while a provider records what it accepted and reported delivering. Comparing selected records can reveal agreement, conflicts or gaps without requiring a central authority to hold everyone’s data.

Participants retain control of their information and can disclose additional details when evaluating outcomes.

Both parties need not install the software for one to benefit, although compatible implementations allow them to exchange evidence. Verification can occur offline and is free.

The system uses Agent Action Capsules, an open format for signed records, and cryptographic checkpoints that make subsequent changes detectable. A free public witness establishes that a checkpoint existed by a particular time.

Preserving a record doesn’t establish that its contents are accurate or that every action was captured, Mih said. EvidenceBook verifies that reported events actually occurred but can’t verify that an agreement was satisfied.

That limitation carries through to Countersign. The service checks specified cryptographic and structural properties of an evidence bundle, signs its findings and records them in its own transparency log. It does not certify that claimed outcomes occurred or that the evidence is complete.

Terms Packs address whether recorded work met specified requirements according to regulatory or business rules. Packs define evaluation criteria and methods, with results identifying requirements that were met, missed or could not be evaluated.

An initial pack covers agent-related obligations under the European Union’s AI Act. Action State recommends that governance, risk and compliance teams customize packs to reflect their requirements. The packs do not certify legal compliance.

Human review is critical to the process. Mih described a workflow in which AI judges assess agent records daily, with employees periodically reviewing samples to compare human and automated judgments. The system also preserves evaluation results, allowing auditors to examine the underlying evidence.

Early integrations include Mesh-LLM, an open-source distributed inference network where participating nodes keep signed records and can compare exchanges with peers. The approach could help participants spot changes to a counterparty’s recorded claims about models or configurations.

Project Nanda, an open-source initiative aimed at building infrastructure and protocols for AI agents, has also used Agent Action Capsules in transactions between agents discovered through separate registries. Each side retained a sealed bilateral record. For personal agents, Action State’s Everyday Rules integrates with Muse to evaluate proposed actions against an owner’s requirements.

EvidenceBook is available under the Apache 2.0 license. Its underlying specifications build on published Internet Engineering Task Force standards, but Action State’s own specifications remain individual Internet-Drafts.

“It’s not an RFC standard or adopted by a working group yet,” Mih said, referring to the IETF’s request for comments procedure.

Action State plans to seek neutral foundation governance for the specifications and implementations. The company expects to earn revenue from independent verification services.

“Countersign is currently free,” Mih said. “However, once there is a relying party that accepts the Countersigns, then we will start to charge for that.”

Image: SiliconANGLE/Microsoft Designer