ActionState EvidenceBook: Agent Work Anyone Can Independently Check
Action State Group announced an open evidence layer for AI agents at AGNTCon + MCPCon North America on 22 October 2026. EvidenceBook, Apache-2.0, lets an agent keep a tamper-evident record of what it did and what it agreed, checkable by anyone without trusting its operator; Terms Packs define the rules used to assess that work, and Countersign adds an independently operated second check
Official Links
Features
- ✓ EvidenceBook gives each party its own signed, tamper-evident history: the buyer agent records what it asked for and paid, the provider agent records what it accepted and delivered
- ✓ Comparing the two records surfaces agreement, conflicts or gaps instead of relying on one side's logs
- ✓ Terms Packs evaluate recorded work rule by rule against business or regulatory requirements, returning met, missed or could-not-evaluate; the first pack covers agent-related obligations under the EU AI Act
- ✓ Countersign is available today as an independently operated verification service, so the check does not depend on the party being checked — no agent grading its own homework
- ✓ Apache-2.0 open source on an open-standards path, alongside work on hardware-attested agent governance records (TRACE): your data stays with you and only the proof travels
How to Use
- Decide what you need to prove: that something happened, or that it complied with a specific rule. The first needs EvidenceBook records; the second needs Terms Packs
- Instrument the action boundary: route tool calls through a gateway or proxy and sign records there, rather than having the agent write its own log lines in code
- Run the most relevant Terms Pack against historical tickets as a backtest and watch the could-not-evaluate rate — a high rate means your record granularity is too coarse
- For cross-company work, add Countersign verification and archive the evidence bundle with the verdict so auditors get both in one place
FAQ
What core problem does EvidenceBook solve?
The stated problem: agents now act on behalf of people and interact with each other across companies and platforms, so there must be a way to independently verify what an agent did, regardless of who operates it or which systems it uses. EvidenceBook is an Apache-2.0 open evidence layer that lets an agent keep a tamper-evident record of what it did and agreed, checkable by anyone without trusting the operator. Mazin Gilbert, executive director of the Agentic AI Foundation, publicly backed the vendor-neutral evidence layer. Site at https://actionstate.ai with the release carried by EINPresswire
How are Terms Packs different from an ordinary rules engine?
The object being evaluated is the recorded work. Terms Packs compare records against business or regulatory requirements item by item and return met, missed or could-not-evaluate rather than one aggregate score, with the first pack covering agent obligations under the EU AI Act. The public material is also explicit about the limit: it can only judge behaviour that reached the record, so behaviour never recorded cannot be assessed
What is Countersign, and is it mandatory?
Countersign is an available Action State product positioned as a second, independently operated verification of the evidence, so the verdict does not rest on the transacting parties. Whether you need it depends on context: internal audit may be fine with EvidenceBook alone, while cross-company transactions or evidence shown to regulators and customers are where independent verification earns its keep. Confirm capabilities and pricing on the vendor site
Where do deployments go wrong?
At the recording boundary. If the agent writes its own logs it can also omit or rewrite them, which voids the evidence. Route tool calls through a gateway, sign and persist there, and put storage where the agent's credentials cannot write or delete. The second failure is granularity: too coarse and Terms Packs return could-not-evaluate everywhere; too fine and cost and volume climb — work backwards from the audit question you must answer
How does this relate to observability platforms?
Observability answers how the system is running — latency, error rates, traces. An evidence layer answers who authorised an action, what the outcome was, and whether a third party can verify it. They coexist: a trace tells you which step was slow, an evidence bundle tells you whether that step was authorised and whether you can show it to an auditor. Where payments, acting for customers or regulated processes are involved, the evidence layer is usually not optional