Your agents touch your most sensitive data. Their limits should come from it.
Agentry sets every AI agent's limits from the classification of the data it uses. You label your data once; each agent's tools, actions and approvals are worked out from that label, and every decision is recorded.
Book a working demonstration See how it works
Components running today · the integrated platform is in development
An agent can never be given more freedom than its most sensitive data allows
The data it uses running today
Each dataset carries a class set by its owner. An agent's class is the highest class among the datasets a person granted it, and a versioned matrix turns that class into an answer for every tool and action: allowed, needs a person, or refused. Grant one restricted dataset and the agent's send tools close at once. Grant nothing and it can do nothing.
The data it produces in development
We are extending the same rule to what agents generate. An output will be classified at least as high as the data that produced it, higher when records are combined, never lowered without a named person, and that class will decide where it may go. A summary of a restricted record stays restricted.
No permission is typed by hand. Change the matrix once and every agent is recomputed, with the change in the record.
One of your agents does something it shouldn't. The board asks four questions.
Who owns that agent?
Not the team. The person.
Who allowed it to act?
On what rule, recorded when it happened.
Can we show what it did?
From evidence, not from recollection.
Do others share the weakness?
And what single change would clear it.
If answering those would take a week of emails and log searches, we should talk.
Six parts, in the order the work happens
Each one is a screen in the console, and each leaves a record.
Access follows the data's classification
You label your data once. Each agent's permitted tools and actions are worked out from the class of the data it was granted, so nobody types a permission by hand.
Every agent gets an identity and limits
A named owner, the class of data it may see, the most it may do, the tools it may call, and what it may spend.
Governed from idea to retirement
Seven stages with gates: a recorded mandate, a baseline, tests, an approval by someone other than the builder, and re-approval whenever the model or tools change.
Every action gets a decision
Permit, deny, or hold for a person — with the rule and the reason recorded. An agent can never approve its own work.
Every call leaves evidence
Agents connect through a gateway using the widely adopted chat-completions interface, so most need only a changed address. Each reply carries its decision, the controls applied, and fingerprints of what was sent and returned.
Evidence becomes a ranked list
Risks are grouped by the control that clears them, so one change can lift many agents — and next week's evidence shows whether it worked.
Five questions we score every run against
| Responsibility | Who owns this agent, and who approved what it did? |
|---|---|
| Auditability | Can we show what happened, from evidence rather than recollection? |
| Interpretability | Why was it done, in words a person can understand? |
| Dependability | Does the result hold every time? |
| Traceability | Where did it come from? |
Missing evidence caps a score rather than nudging it, because a score you cannot evidence is not a score. Every figure says where it came from: measured, sampled, assumed, or not computable with the reason.
Four lines of work, and you can start on any one of them
Most organisations do not need all four at once. They need the one that is currently stopping them.
Agents Strategy
Where agents belong, where they do not, and what the board is being asked to accept.
before you build
Agents Delivery
Getting agents built and released through gates, rather than released because someone was confident.
while you build
Agents Assurance
Evidence of what each agent did, scored, so the answer survives an auditor.
once they run
Agents Capability
Your own people running this without us, with the standards written down.
so it outlasts us
Three people, three first screens
If you own the risk
The estate in one screen: how many agents, where they are, which practices are behind, what got through that shouldn't have, whether last month's fix worked — and one number that must stay zero.
CISO · head of AI · risk and audit
If you run a team that uses agents
Your day is a queue: decisions the agents may not take alone, outputs to judge, exceptions to chase — each with its context and its consequence.
operations · service management
If you build and own agents
What to finish next, what a gate is asking for, whether your agent is running, and what people said about its work — without becoming a governance specialist to find out.
agent owners · delivery teams
See it working
Thirty minutes, on screen. We will show you what runs today, say what is still being built, and answer what we cannot do yet.
Book a working demonstration