You cannot control what nobody wrote down.
Most organizations find out during an audit that nobody has a current list of the AI they are running. Here the list writes itself while policy is enforced: an agent enters the register on its first call.
Rules only, so changing one needs no release.
What is on the record for each agent
| What it is | name, owner, business purpose |
| What it sits in front of | the model, the provider, the endpoint |
| What it may call | tools, endpoints, and the limits on each |
| What governs it | the ruleset and the version in force |
| What it has done | every decision, with the rule that produced it |
Nobody types an agent into it. The first time an agent calls, its row is written in the same transaction that records the call. The row is written even when policy refuses that call: refusing an agent is not a reason to leave it out of the inventory. You declare an owner and a risk level on top of a row the traffic already created, so the register cannot go out of date the way a spreadsheet does.
Documentation you write against
The declarations, the rule versions and the decision history are the raw material for a technical file: what the system does, what controls sit on it, what it decided and on which version. Export it and write.
We do not decide your system's risk tier. That is a legal decision about how you use the system, made by you, usually with your lawyers. Any vendor offering to make it for you is selling something they cannot deliver. See our read of the EU AI Act for where that line sits.