Contain the agent, sheet 05 of 13
Whether an agent acts as a user or as itself, how it gets credentials, and how to keep real keys out of its hands.
Broker credentials instead of handing out keys. The agent should never hold the real secret.
← → move to the next sheet over, Esc zooms out.
When an agent does something, it does it as somebody. Either it borrows a person's identity and acts on their behalf, or it has an identity of its own. Most teams haven't decided, which means it's quietly using whichever API key was lying around.
It's the difference between lending a contractor your house key and giving them a lockbox code that only opens the garage, only this week, and logs every visit. Both get the job done. Only one survives the contractor's laptop getting stolen.
The identity vendors have caught up. There's a well-trodden path for an agent acting as a signed-in user, and agents can now have identities of their own. The hard case is still the agent running at 3 a.m. with nobody signed in, and that one is mostly up to you.
Every link opened on 11 Nov 2026. The numbers match the squares on the map.
Standard and product, Okta
An OAuth extension that lets a company's identity provider govern which apps and agents may connect to which, including connections to MCP servers.
The catch. The early-adopter list is Okta's own. Check that the tools you use actually support it.
Product, Auth0
Stores the access and refresh tokens of external providers so an app or agent can call their APIs on a user's behalf, without holding the user's credentials itself.
The catch. Delegated access is only as narrow as the scopes you request. Ask for less.
Product, Microsoft
Identities for AI agents, created from blueprints. With an Agent 365 license, Conditional Access, identity protection, governance and audit logs of their own.
The catch. The identities are open to every Entra customer. The security features for agents need Microsoft Agent 365, included in Microsoft 365 E7 or sold as an add-on.
Product
Runtime access control for agents and workloads: it enforces policy, keeps downstream credentials out of the agent, audits access, and can revoke it with a kill switch.
The catch. Another control plane to run. Worth it once you have more agents than you can name.
Product
One identity layer for people, machines and AI agents across infrastructure.
The catch. Infrastructure access first. It's broader than the agent problem, which is a plus and a project.
Standard, open source
The open standard for workload identity: a verifiable identity document a piece of software gets from where it runs, instead of a shared secret.
The catch. A building block, not a product. Most teams meet it inside a platform rather than running it alone.