The Agents page shows you activity. Registry answers a different question — the one auditors ask first: show me your inventory. Every agent in your estate, classified by risk, approved or pending, with a named owner, and a register you can hand over as a document. It's the difference between having agents and governing them.
Discovered automatically, governed explicitly
The list itself builds from the stream — any agent that acts is in it, so the inventory can't quietly fall out of date. What the stream can't supply is judgement, and that's what Registry adds on top: for each agent, a risk tier under the EU AI Act's classification scheme, an approval status, and an assigned operator — a human name accountable for the agent's behaviour.
The named owner is half the value of the whole page. "Whose agent is this?" stops being a Slack archaeology exercise, and the register stops being a list of software and becomes a map of accountability — which is what a governance reviewer is actually checking for. An agent everyone uses and nobody owns is the organisational shape most incidents grow out of; a register makes that shape visible before the incident volunteers to.
The classification workflow
New agents arrive unreviewed, and the working rhythm is simple: filter to Unreviewed, classify each one, approve before production. Two details show the register is built for real audits rather than appearances. Classifying an agent High or Prohibited requires a written justification — the EU AI Act's Article 11 documentation expects the reasoning, not just the label, and "we'll write it up later" is exactly the debt that surfaces during an assessment. And agents left sitting in Pending review are flagged in your Evidence Packs, so an unfinished inventory is visible instead of cosmetically complete. That flag has teeth by design: it means the shortcut of ignoring the review queue shows up in the very document you hand an auditor, which is the right incentive pointed in the right direction.
Classification changes are themselves recorded
Who moved this agent from High to Limited, and when, and on what justification? The register keeps the audit trail of every classification change. This closes a loop auditors genuinely probe: an inventory whose labels can be silently rewritten is barely better than no inventory. Here, governing the agents produces its own governed record — the same principle as the rest of the product, applied to the governance layer itself.
The register as a document
When the audit cycle comes around, Registry generates the Register PDF: a snapshot of every agent with its tier, approval state, operator, and event volume. That document is the artefact an external auditor reads for technical-documentation compliance — produced from the live inventory in one click, not assembled in a spreadsheet the week before.
The difference shows on audit day. A spreadsheet register invites the follow-up questions that eat afternoons: is this current, who maintains it, how do you know nothing's missing? A register generated from the recorded stream answers all three in its construction — it's current because it was generated today, maintained by nobody because it maintains itself, and complete because appearing in it is a side effect of acting at all. The auditor's questions move on to something that deserves them.
Registry is where the census from Agents becomes governance: activity plus classification, ownership, and the paper trail. Questions about the classification scheme? Ask us.