Use cases

Where a sealed room for AI agents pays off.

Anywhere an agent can move money, read records or send data, and someone will later ask you to prove what it did.

A made-up example. Different kinds of work share one sealed room, one guarded door and one logbook.

Pick yours

Six kinds of work, one sealed room.

Each one is a workflow we are piloting or scoping with design partners. Find the one closest to yours.

These are the workflows we are piloting with design partners, not customer results.

Banks and payments

Agents that match payments, chase missing files and line up transfers.

The work: Read invoices and bank files, match them, and queue each payment for a person to approve.
The risk: A payment sent twice after a crash, or account data sent to a service nobody approved.
What your auditor sees: Every step in the logbook, a restart that repeated nothing, and a signed record of what ran and where it connected.

Insurance

Agents that read claim files, check them against the policy and draft a decision.

The work: Pull the claim, the policy and the photos, check them against the rules, and draft a decision for a handler.
The risk: Customer data reaching a model or a site that you did not choose.
What your auditor sees: The list of places each agent could reach, every call it made, and every call that was turned back.

Healthcare

Agents that summarise records and prepare referrals inside the hospital network.

The work: Read records on your own network and write short summaries for doctors and nurses.
The risk: Patient data leaving the network on its way to an outside service.
What your auditor sees: Patient data could reach only the systems you named. Every try to go somewhere else was blocked and written down.

Public sector

Agents that sort cases and prepare files on your own hardware.

The work: Sort new cases, gather the papers and prepare a file for an officer to decide.
The risk: Calls to outside services on a network that must stay sovereign, or offline.
What your auditor sees: The machine runs on your own hardware, with no outside calls unless you approve them.

Platform teams

Give every business unit one approved place to run agents.

The work: Offer teams a ready, approved way to run agents, so each new agent does not start a new security review.
The risk: Ten teams, ten setups, and no common evidence when the auditor asks.
What your auditor sees: One sealed image, one approved list per agent, and the same kind of record for all of them.

Software companies

Run your product's agents inside your customer's own cloud.

The work: Ship your agent product into a bank, insurer or hospital, running in their account.
The risk: Their risk team will not take your word for what your agent does.
What your auditor sees: Proof from the machine itself, which their auditor checks without having to trust you.

How it works

You approve where data may go, the agents work in the room, and your auditor checks the proof.

List the systems and models each agent may reach. Everything else is turned back at the one door.

Each agent is one simple file. Every action goes in the logbook before it happens, so a crash does not repeat it.

The machine signs what it started with and where it connected. Your auditor checks it on their own computer.

0 duplicatesand 0 lost across 5 forced-crash tests. See the test

Measured 2026-10-14 on the image built from tag phase-7, median of 3 runs where run three times, 2013 Intel i7-4770 under KVM.

Tell us about it. Design partners use it free through 1.0 and help decide what the proof must show.