Choosing an agent identity platform: 20 questions

How should I use these questions?

Send them to each vendor before a demo. Then ask the vendor to show, not tell, the answer to each one in the demo. Mark an answer as shown, described or missing.

The questions follow the seven criteria in our scoring method. The weights there reflect our view. Your own weights may differ. The best for pages show how rankings change when weights do.

Ownership

Question one. How does an agent get registered, and what fields are required? Look for identifier, owner, purpose and lifecycle state as a minimum.

Question two. How do you find the right owner for an agent you discovered? Ask how the product joins a person's accounts across identity providers, directories and HR.

Question three. What happens to an agent when its owner leaves the company? You want an automatic pause or reassignment, not just a stale field.

Credentials

Question four. Does the agent ever hold a credential for the tools it calls? The strongest answer is no, because a broker or gateway attaches it.

Question five. How does the agent prove its own identity without a stored secret? Expect workload identity, SPIFFE or a local broker.

Question six. How long do issued credentials live, and who sets the lifetime? Expiry should not depend on anyone remembering to revoke.

Question seven. Are tokens sender-constrained with DPoP or mTLS? This matters if tokens pass through many hops.

Delegation

Question eight. When an agent acts for a person, does it stay within that person's permissions on every call? Ask to see a denied call when the person lacks access.

Question nine. Show me a log entry for a delegated call. It should name the agent, the person and the decision.

Question ten. How are agent-to-agent chains recorded, and can I limit their depth? A chain should trace back to a human.

Runtime control

Question eleven. Is every tool call checked against policy at the moment it runs? Sign-in checks alone are not enough for agents.

Question twelve. Where does enforcement happen: a gateway, a proxy, an SDK, or nowhere inline? Detection alone does not stop an action.

Question thirteen. Can a risky action pause for a named person to approve it? Ask how the approver is notified and how they confirm.

Question fourteen. Can an agent ever approve its own action, or another agent's? The answer should be a firm no.

Question fifteen. Can I test a new policy against real past sessions before it goes live? NewCore describes replaying policies against recent real sessions. That reduces surprises.

Discovery

Question sixteen. How do you find agents nobody registered? Ask for the list of sources: identity providers, SaaS, code, CI/CD, endpoints, AI platforms.

Question seventeen. Do you trace access paths outside SSO? Agents often use personal tokens that never touch the identity provider.

Standards

Question eighteen. Which of these do you support today: OAuth 2.1, MCP authorization, Cross App Access or ID-JAG, CIMD, SPIFFE, DPoP, mTLS? Ask for documentation, not a roadmap slide. Our standards tracker lists what each scored platform names in public material.

Buying clarity

Question nineteen. Is there a published price or a free tier? Aembit publishes per-agent pricing and a free Starter tier. Descope has a Free Forever tier. Most others are quote-based.

Question twenty. How do we buy: direct, through a cloud marketplace, or inside a larger suite? A marketplace listing can shorten procurement. NewCore sells through AWS Marketplace. SailPoint Entro is sold within SailPoint suites.

What about the identity provider we already run?

Ask every vendor one extra question. Does your product run alongside our identity provider, or does it try to replace part of it?

An agent identity layer runs alongside your identity provider, never instead of it. People keep signing in where they do today. The layer adds discovery, per-call checks and approvals for agents.

NewCore states that your current identity platform keeps handling employee sign-in. Aembit states it does not replace the workforce identity provider. Silverfort enforces SSO through your identity provider.

How do I score the answers?

Use the same scale we use, zero to ten per criterion. A shown answer earns full marks. A described answer earns partial marks. A missing answer earns little or nothing.

Then apply your weights. If you run Okta and want to keep it, weigh runtime control higher. If you are cleaning up shadow agents, weigh discovery higher. Our keep Okta and shadow agent cleanup pages show those weightings.

Which answers should end a conversation?

A few answers should make you pause.

• The agent stores a long-lived key for production tools.
• There is no way to block an action inline, only alerts after the fact.
• Owners are a free-text field with no link to the directory.
• The vendor asks you to move employee sign-in to their product.

None of these is always fatal. Each one deserves a hard follow-up.

A downloadable version

These questions are also available as a spreadsheet you can fill in per vendor. Download the evaluation checklist. The current figures for every scored platform are in the standings file.

How long should the evaluation take?

Plan for a few weeks, not months. Spend the first week on discovery against one business unit. Spend the second on credentials and delegation with two or three real agents. Use the third for runtime control and approvals on one risky action.

Keep the same agents and the same test cases for every vendor. That makes the answers comparable. Record each demo, so the scores can be checked later by someone who was not in the room.

What should the shortlist look like?

Most teams end up with two or three vendors. Usually one is strong at discovery and ownership. Another is strong at credentials and standards. A third covers runtime control and approvals.

Some products cover all three reasonably well. That is what the overall standings try to capture. But a buyer with a narrow need may do better with a specialist. Run the demos, score the answers and let your weights decide.