Action State makes open tools that record the work of AI agents

Action State made EvidenceBook, an open-source toolkit that helps companies check the work of AI agents. The agent of one company can hire the agent of a different company to do a task. Steven Mih is the founder of the company. The founder gives this view: the records must be free to check and must go with the work. The company gives no test numbers.

How it works

The buyer records the task and the payment. The provider records the work that it accepted and delivered. Then the 2 sides compare selected records. They see which parts agree, which are different and which have no proof. A Terms Pack is a group of rules that the system applies to the records. Countersign makes the check as a third party.

The proof

The source gives no test results. In a live deal room, agents from 2 registries closed deals and each held the same record. Mesh-LLM, an open-source network, also uses the records between nodes.

What is not known

A check shows that the records did not change. It does not show that the work was done. Countersign does not show that all work was recorded. The specifications are drafts of one internet group, not accepted standards.

Insight: Agent trust is built in different layers

On 14 October 2026, 4 companies showed open identity checks for agents. Researchers in Spain made a study. They found that an identity check does not show that one instruction follows policy. EvidenceBook adds a new layer: a record of the work. Each layer has a different builder, and this can be a race to set the open standard for agent trust. Action State plans to propose its drafts to a neutral foundation.