Agents misread tasks, skip steps, and report successes that didn’t happen. EvidenceBook keeps a tamper-evident record of what your agent did and agreed, so the other side can check it without trusting you.
It makes things up.
Reported 44 invoices reconciled. 41 were.
It covers its tracks.
The retry that failed isn’t in the summary.
It misunderstands.
Asked to pause the campaign. It deleted it.
It tries too hard.
Told to cut costs. It cancelled the backups.
It misses what matters.
Shipped on time. Skipped the security review.
It makes mistakes.
Paid the right vendor twice.
It gets it plain wrong.
Booked the flight for March, not May.
Not because agents lie. Because the only account is the agent’s own. A second record, written at the time by the other side, is what shows where they differ.
Models, tools, payment rails, and identity are all being built. Evidence of what an agent actually did is the missing layer.
“Discovery tells you an agent exists. It cannot tell you what that agent did after you trusted it.”
Trust asks you to believe the agent. Evidence lets you check it: sealed as it happens, so any later change shows, and readable by the other side.
One company’s agent hires another’s. The buyer records what it requested and paid. The provider records what it accepted and delivered. Exchange selected evidence and see what agrees, what conflicts, and what remains unconfirmed.
Verification shows what the records say and that they haven’t changed. It doesn’t show that the events they describe occurred.
Called the payment API for order 4471.
Received and executed tx_9c1e for order 4471.
You delegate so you don’t have to watch every step. The record shows what it did while you weren’t looking.
Your customer never sees your logs. They see evidence they can check, and only the parts they’re allowed to.
Two agents, two companies, one transaction. Each side keeps its own record, and either benefits even if the other doesn’t.
Stand-in. The independently checked sample report is not published yet. Until it is, order 4471 below is an illustration, not a sealed record.
An agent does something that matters.
Each boundary it touches records what it saw.
Observations go into append-only, tamper-evident history.
The other side asks for the evidence it is allowed to see.
They check it themselves, offline, without trusting us.
A Terms Pack runs over the records: met, not met, or could not be evaluated.
Your host decides whether the action proceeds.
A Terms Pack is a versioned list of requirements, run over the records. Some are easy to check. Others need judgment, and say who judged.
A fact or rule over the record. Anyone who runs it gets the same answer.
A pinned AI model or a named person decides. The verdict records which, and the judging steps are readable.
Packs don’t certify compliance or score trust. Anyone can publish one. Feature coverage varies by integration.
Usage graphs don’t renew contracts. Run the Outcomes Pack over the work: outcomes confirmed by the customer, judged against their bar, with the misses shown too. That is how ROI gets believed.
Which model ran, what policy governed it, who authorized it, and what changed, written down at the time. The EU AI Act Obligations Pack shows the evidence behind each finding. It doesn’t certify compliance.
Before an agent moves money, signs, or deploys, see every prior step that led there. Everyday Rules checks the action before it happens. Countersign adds an independent second signature where the stakes warrant it.
Either side can verify the other’s records today, for free. But in a dispute, both sides have an interest. Countersign runs the checks as a party independent of both, signs exactly what it checked and found, and logs the result in its own transparency log.
Version 1 runs five checks. Each result stands alone, never combined into a score, a grade, or a single pass/fail.
Stand-in. The independently checked sample report is not published yet. Until it is, this finding is an illustration: bundle 3a70d2e4 and log entry 71fa0c3e are not published records.
“For that to work, we need a way to independently verify what agents are doing, regardless of who operates them or which systems they use.”
Each negotiation produced a sealed bilateral record both agents hold, alongside each side’s own account. The integration merged into the NANDAtown codebase in July 2026.
“Watching agents from two different registries close deals and walk away with the same cryptographic record … was the missing half of the open agent economy running live.”
An open-source distributed inference network with no central evidence authority. Each node verifies its own records. When both opt in, peers exchange evidence to compare what each observed.
“Our goal is to help Mesh-LLM nodes collaborate with peers they’ve never met.”
Customers, partners, and auditors verify the evidence themselves. Nothing is published by default. Each party sees only what it is allowed to.
See a sample report →Stand-in. The independently checked sample report is not published yet. Until it is, the figures here are an illustration, not a sealed record.
Open-source libraries with MCP bindings today and an A2A binding in preview, inside your existing agent harness. Seal each consequential action, keep it in a tamper-evident log, and hand the other side a bundle it can verify.
Read the capsulectl docs →A single capsule is checked against your profile’s trusted keys; inclusion in the log is not checked here. capsulectl verify --bundle checks it, for every record in a bundle.
A plain-text index for agents: what installing Action State puts on a machine, and every SHA-256.
GET /llms.txt →# Action State > actionstate.ai is the company site of Action State Group. https://actionstate.ai/install lists what installing Action State Receipts puts on a machine (capsulectl, the deal skill and the Action State plugin), every SHA-256, and the install skill a person gives their agent. The same page, as Markdown: https://actionstate.ai/install.md ## Install - [Install Action State Receipts](https://actionstate.ai/install.md): for the person. What gets installed, the versions and every SHA-256, the sites the install reaches, and the install skill (a SKILL.md with its SHA-256) that holds the steps.
Anyone can produce compatible evidence, define requirements, and verify records. The drafts ship with conformance vectors and build on SCITT (RFC 9943) and COSE Receipts (RFC 9942).
Action State is a Silver Member of the Agentic AI Foundation and plans to propose the specifications for neutral foundation governance. Membership does not constitute acceptance of the proposed projects.
Built on the same open specifications, so others can offer compatible services. Feature coverage varies by integration.
EvidenceBook has not yet passed its own interoperability test: an outside developer exchanging evidence with it without using our code. Until it does, we make no interoperability claim for it.
You stop asking to be trusted. They stop having to.