Evaluate agent identity products on six things: whether each agent gets its own identity, how credentials are scoped and expire, how delegation from a user is represented, how fast access can be revoked, how actions are attributed in logs, and whether the product works with the systems you already run.
Why shared accounts fail
Many agents today run on a shared service account or a developer's personal token. That makes audit trails useless: every action looks like it came from the same person. It also means one leaked token exposes everything that account can reach. The MCP security guidance notes that token passthrough breaks accountability and that broad scopes enlarge the damage from a stolen token. Identity for agents is about fixing both.
The six checks
-
Unique identity. Can you create one per agent or per task, and see it listed in an inventory?
-
Scoped, short lived credentials. Can scopes be narrow, and do credentials expire without manual work?
-
Delegation. When an agent acts for a user, does the record show both the agent and the user, or does it impersonate?
-
Revocation. Can you cut off one agent in minutes without breaking others?
-
Attribution. Does every tool call in the log carry the agent identity and the policy decision?
-
Integration. Does it work with your identity provider, secrets store and the tools agents use?
A practical test
Create two test agents with different scopes. Ask each to perform an action outside its scope and confirm it fails and is logged. Revoke one mid task and confirm it stops. Then ask an auditor, or a colleague playing one, to reconstruct who did what from the logs alone. If they cannot, the identity story is incomplete regardless of the feature list.
When you may not need a product yet
If you run one or two agents with tightly limited tokens issued by your existing identity system, careful scoping and logging may carry you for a while. Dedicated agent identity matters when agents multiply, when they act across systems, or when you need to show evidence to a customer or auditor.
Ask one more question about lifecycle. Agents are created quickly and forgotten quickly. What happens to an identity when its project ends? Look for expiry by default, ownership fields that name a responsible human, and reports on identities that have not been used recently. Orphaned agent credentials are the same problem as orphaned service accounts, and they tend to be found only after something goes wrong. A tool that makes stale identities visible is doing real work.
Where Safeguard fits
Safeguard includes agent identity, MCP Guardrail and asset discovery for agents and MCP servers.
Quick check
Is an API key per agent enough?
It is a start, since it gives separate attribution. Add scoping, expiry and revocation to make it a real identity control.
Sources
Last reviewed: 19 October 2026.
