Startup Action State launches tools for verifying interactions between AI agents

Action State Group Inc. is trying to solve a very specific problem: once AI agents start buying, selling and handing off work across company lines, whose record do you trust?

The startup has launched a set of open-source tools meant to let businesses verify those interactions without leaning on a central referee. The bundle combines EvidenceBook, which keeps tamper-evident records; Terms Packs, which judge agent behavior against set rules; and Countersign, an independent verification service.

Steven Mih, Action State’s founder and chief executive, says the key difference from observability tools is scope. Those systems usually show what happened inside one company’s own walls. Action State’s approach is aimed at exchanges where two parties each need proof of what they requested, accepted and delivered.

EvidenceBook gives each side its own signed transaction history. A buyer can log what it asked for and paid. A provider can log what it accepted and what it says it delivered. Those records can then be compared for agreement, conflicts or gaps. Both sides do not have to install the software for one side to benefit, and verification can happen offline at no cost.

But the company is also drawing a line around what the tools do not prove. Preserving a record does not mean the contents are true, and it does not guarantee every action was captured. Countersign checks the structure and cryptographic properties of an evidence bundle, then signs off on that check in its own transparency log. It does not certify that the claimed outcome actually happened or that the evidence is complete.

Terms Packs take a different bite at the same problem by testing whether recorded work met requirements defined by regulatory or business rules. The first pack deals with obligations tied to the European Union’s AI Act. Action State says governance, risk and compliance teams should tailor packs to their own needs, and that human review stays critical. The company is already talking about daily AI reviews with periodic employee spot checks.

There are early integrations with Mesh-LLM and Project Nanda, both open-source efforts in the AI-agent world. For personal agents, Everyday Rules ties into Muse to judge proposed actions against an owner’s requirements. EvidenceBook is under the Apache 2.0 license, while Action State’s own specifications are still individual Internet-Drafts. Mih says the company wants neutral foundation governance eventually, and expects revenue from verification services once someone is willing to rely on Countersign.