Home / Blog

AI agent inventory: what to track for every agent

3 min read · Published 17 October 2026

Before you can offboard, audit or limit an agent, you have to know it exists. Most companies cannot say how many agents they run. A simple list fixes more than you would expect.

Short answer

Keep one list of agents with, at minimum, a name, an owner, a purpose, a status, what the agent can reach, its credentials and their expiry dates, when it was created and when it was last used. The IETF SCIM draft covers the first four. You will track the rest yourself.

Why the list comes first

Almost every control you want depends on it. You cannot offboard an agent you do not know about. You cannot answer an auditor's question about who owns it. You cannot shrink its access if you never wrote down what it has.

The OWASP Non-Human Identities Top 10 has no item that is literally called inventory. But Improper Offboarding, Overprivileged NHI and Long-Lived Secrets all assume you know the identity exists. The list is the quiet requirement underneath the others.

What to record

The last column shows whether the field exists in the core schema of the June 2026 SCIM agent draft, revision 00, or in Oracle's optional governance extension.

FieldWhy you want itIn the draft?
NamePeople need to refer to it.Yes, displayName
Unique IDOne stable key to use in every system.Yes, agentUserName
OwnerSomeone must answer for it, and be asked when it misbehaves.Yes, owners
PurposeWithout it, nobody dares to switch the agent off.Partly, description
StatusActive, paused or retired.Yes, active (on or off only). The governance extension adds five states.
AccessWhich systems and data it can reach.No
Credentials and expiryTokens, keys and certificates, and when each one dies.No. The governance extension has a credentialDiscoveryUri.
Valid untilA date after which the agent should stop working by itself.Governance extension: validTo
Human oversightWhether a person approves what the agent does.Governance extension: autonomyLevel
Created and last usedFinds agents that nobody uses any more.No

WorkOS notes that the earlier Okta proposal had more, including certificates and a last-accessed timestamp, but on the application resource rather than on the agent. Whatever the standard ends up including, keep these fields in your own list. They are cheap to record on the day you create the agent and painful to reconstruct later.

Where to find the agents you forgot

Your first list will be incomplete. That is normal. Look in these places:

Keeping it alive

A list that is right once and wrong in six months is worse than no list, because people trust it. Three habits help:

  1. Make registration part of creating an agent. No record, no credentials.
  2. Have owners confirm their agents on a schedule. Quarterly is a reasonable start.
  3. Flag orphans automatically. An agent whose owner is inactive should raise an alert. The offboarding checklist says what to do next.

Where SCIM comes in

Once the list exists, SCIM is the way to keep apps in sync with it. The directory holds the record. Each connected app receives the same identity, and a change to the status reaches all of them. That is the idea behind the agent extension that the IETF is drafting, and you can try the effect yourself in the demo on the home page.

The standard is still a draft, so do not wait for it. A spreadsheet with the fields above is a perfectly good start.

Sources

Own the name for this category.

agentscim.com is for sale. If you build identity, security or agent tooling, write to me. Serious offers get a reply within 24 hours.

Make an offer

Keep reading

This is an independent page. It is not affiliated with the IETF, Okta or Microsoft. Standards and drafts change, so check the linked sources for the current state.