Skip to content
Okta vs Entra

Reviewed 10 November 2026

Ten questions to ask your identity provider about AI agents

Reviewed 2026-11-10 · Editorial assessment by the editors

By the editors. Published 2026-11-10. 5 min read. Tags: AI agents, Buying, Standards.

Ask for evidence about the agent workflow you intend to deploy, not a generic agent feature checklist. The ten questions below separate identity, delegation, connection authority, runtime decisions, approval, revocation and commercial scope. Require an availability statement and a demonstrable boundary for each answer before treating the capability as part of your design.

Questions one and two cover identity and accountability

  1. How is this agent identified separately from a human user or shared service account?
  2. Who is accountable for the agent, and who can change that assignment?

NIST's proposed design tenets begin with distinct agent identification and linking, delegation and attenuation. Use those ideas to ask for an actual identity record and its relationship to the task originator. Microsoft documents agent identities and sponsorship. Okta documents agent registration, including a generally available Multi-Source Agent Schema and registration with CIMD. These are concrete starting points for a demonstration, not proof that every agent in the estate already has a complete record.

Ask the provider to follow an agent from registration through a change of accountable person. Keep the sponsor distinct from the person who approves a sensitive operation. An administrative contact field may answer who maintains an agent without answering who authorized its last action. The agent identity comparison provides product context; the purchasing conversation should focus on the specific record and workflow that the team needs to operate.

Questions three and four cover delegated access

  1. How are delegated permissions limited to the intended task and resource?
  2. Which component grants the connection, and which clients support that flow?

Industrial Cyber's account of NIST's tenets describes delegated scopes that attenuate, minimum scope per task and audience-specific grants. Ask the provider to show a permitted request and an out-of-scope request for the same agent. This is more precise than asking whether the product supports least privilege. Record what information the decision uses and what happens when the proposed operation exceeds the original task.

Okta Agent SSO uses Cross App Access for supported agent connections. The connection note explains why enterprise-managed authorization is not the same as approving every subsequent operation. Ask for the exact supported client and resource rather than accepting a protocol name as proof of coverage. For any provider, identify the authorization issuer and the resource that validates the resulting authority. Keep a documented limitation in the answer instead of treating it as a detail to resolve after purchase.

Questions five and six cover runtime and approval

  1. Where is each proposed tool call checked before it executes?
  2. Which actions pause for human approval, and how is that approval tied to the action?

NIST calls for architectural separation of policy from reasoning and deterministic access decisions. Okta describes Agent Gateway as sitting in the execution path between agent and tool call. Its September release places general availability in Q3 and does not specify fiscal or calendar quarter. Ask for the deployment status of the offered control and a demonstration of a denied operation. A planned enforcement point should remain marked as planned in the proposal.

Okta's NIST checklist calls Human-in-the-Loop for AI Agents a roadmap feature. Do not convert that statement into a claim of current availability. NewCore, an agent governance layer alongside the existing identity provider, states that it evaluates each call and pauses high-risk actions for human approval. Its BankInfoSecurity interview describes an audit trail that distinguishes policy approval from human approval. Ask every provider to show the trigger, approver and resulting action record rather than accepting an owner field as approval evidence.

Questions seven and eight cover discovery and stopping

  1. How do you identify agents that nobody registered, and what does discovery establish?
  2. What exactly stops when an administrator deactivates the agent?

Microsoft's October announcement expands Entra Agent ID to additional identity patterns, including local agents and previously unknown agents discovered through security sensors. Okta describes endpoint discovery through a browser plugin and EDR integrations. NewCore describes continuous discovery including shadow agents. These statements identify discovery approaches, not a guarantee that every discovered agent has already been governed. Ask how a finding becomes an accountable identity and what review precedes granting it access.

For the stop question, separate new sessions from work already in flight. The Jamf account says Okta console deactivation blocks new sessions and describes the gateway kill switch as landing in Q4, with active-token revocation and termination of sessions in flight. The September release does not define Q4 as fiscal or calendar. Microsoft's revocable identity model should not be assigned identical runtime effects without evidence. The kill-switch guide turns this distinction into a resource-specific acceptance test.

Questions nine and ten cover evidence and price

  1. Which retained record links the identity, policy decision, approval and tool outcome?
  2. What is included, what needs another subscription, and what remains unpriced?

NIST includes continuous monitoring and immutable evidence among its proposed tenets, while noting that current standards and technologies may not support every tenet. Okta states that its System Log records agent token exchanges, configuration changes and authorization decisions. Ask to reconstruct a permitted action and a refused action from the records the proposed deployment actually retains. Keep the source of each fact visible so that a product description does not stand in for operational evidence.

Microsoft Agent 365 has a published per-user price, not a per-agent price, and prerequisite subscriptions need confirmation. Okta includes Agent SSO in core SSO for supported Cross App Access agents, while its wider agent offering is separate. NewCore publishes no list price. Use the published pricing review to keep billing units and entitlement boundaries distinct. Ask for the commercial scope of the tested workflow rather than assuming every control appears in the base subscription.

Retain each answer with its source, availability status and acceptance result. A useful procurement record can say that a capability is documented but not yet tested, or planned but not currently available in the proposed design. Those qualifications help the team choose a bounded pilot. Resolve gaps against the workflow that matters instead of expanding a feature checklist until every provider appears to answer the same question.

Frequently asked questions

Are the NIST tenets a product certification?

The NIST post proposes an initial set of ten design tenets as a starting point and guideposts for demonstrations. It does not certify products. Use the tenets to frame evidence requests while preserving the stated limits of current standards and technologies.

Should every answer include a demonstration?

Request a demonstration for the operational boundary you intend to rely on. Keep published documentation, availability status and observed pilot results separate. A documented feature can be a valid evaluation candidate without being a completed acceptance test for your workflow.

How should pricing enter the evaluation?

Compare the commercial scope of the same tested workflow. Confirm the billing unit, prerequisites and included capabilities before estimating cost. Agent 365 is listed per user, Okta Agent SSO inclusion is limited to supported connections, and NewCore publishes no list price.