Action State launches open-source toolkit for auditing AI agent transactions

Action State Group Inc. unveiled an open-source toolkit aimed at solving a critical trust gap: how to verify the actions taken by artificial intelligence agents when they operate across company lines, making autonomous decisions about purchases, sales and task delegation. The suite brings together three components—EvidenceBook for creating tamper-proof transaction logs, Terms Packs for assessing whether agent conduct met defined standards, and Countersign, a third-party verification service.

The problem Action State is tackling differs fundamentally from what existing observability platforms address. "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," explained Steven Mih, the company's founder and chief executive. Mih previously founded Ahana Cloud Inc., which built a managed offering around the Presto open-source distributed query engine before IBM Corp. acquired it in 2023.

How the system works

EvidenceBook operates by giving each participant their own cryptographically signed record of transactions. A purchasing entity documents what it requested and the amount it paid; a service provider records what it agreed to deliver and what it actually provided. By comparing these records, both sides can identify whether they agree, spot discrepancies or uncover gaps—all without needing a central authority to hold everyone's data.

The system relies on Agent Action Capsules, an open standard for signed records, paired with cryptographic checkpoints that make any later modifications apparent. A public witness service confirms that a checkpoint existed at a specific moment in time, and the verification process can happen offline at no cost.

Participants maintain ownership of their information and decide what additional details to share when assessing results. Notably, both parties do not need to use compatible implementations for one to gain value, though doing so enables them to exchange evidence directly.

What the system can and cannot verify

A critical limitation exists in what EvidenceBook establishes. According to Mih, "Preserving a record doesn't establish that its contents are accurate or that every action was captured." The system can confirm that reported events genuinely took place but cannot validate whether an agreement was actually fulfilled.

Countersign carries the same constraint. This verification service examines specific cryptographic and structural qualities within an evidence bundle, signs off on its findings and logs them in its own transparency ledger. However, it does not certify that the claimed outcomes truly happened or that all relevant evidence was included.

Terms Packs fill a different role by evaluating whether recorded work satisfied specified requirements based on regulatory or business criteria. Each pack outlines evaluation standards and processes, producing results that show which requirements were satisfied, which were not met and which could not be assessed.

The initial Terms Pack addresses obligations that AI agents must follow under the European Union's AI Act. Action State encourages compliance, governance and risk teams to develop customized packs reflecting their own needs. Importantly, these packs do not certify legal compliance.

The role of human oversight

Human judgment remains essential to the entire process. Mih outlined a workflow where AI systems review agent records on a daily basis, with human staff periodically sampling and comparing these automated assessments against their own judgments. The platform preserves all evaluation results, enabling auditors to trace back through the underlying evidence.

Early deployments

Several projects have already begun integrating Action State's technology. Mesh-LLM, an open-source distributed inference platform, has adopted the approach—participating nodes maintain signed records and can validate exchanges with other nodes. This capability helps participants detect if a counterparty has altered its recorded statements about models or system configurations.

Project Nanda, an open-source effort building infrastructure and protocols for AI agents, has employed Agent Action Capsules in agent-to-agent transactions discovered through independent registries. Each participant kept a sealed bilateral record. For individual agents, Action State's Everyday Rules feature works alongside Muse to check proposed actions against user-defined requirements.

Licensing and standards path

EvidenceBook is released under the Apache 2.0 license. Its underlying specifications draw from published Internet Engineering Task Force standards, though Action State's own specifications remain as individual Internet-Drafts. "It's not an RFC standard or adopted by a working group yet," Mih noted, referencing the IETF's request for comments procedure.

Action State intends to pursue neutral foundation governance for both its specifications and implementations. The company plans to generate revenue through its 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."