Skip to content
Okta vs Entra

Reviewed 10 November 2026

Where runtime enforcement sits for Okta and Microsoft

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

By the editors. Published 2026-11-10. 5 min read. Tags: Runtime enforcement, Okta, Microsoft Entra.

Okta describes Agent Gateway in the execution path between an agent and a tool, with policy enforcement and interaction logging. Microsoft's cited Entra material describes Conditional Access at sign-in and token issuance, not per-tool-call checks. Compare those boundaries directly, and keep planned gateway availability separate from evidence of a deployed control.

Locate the decision before comparing products

Runtime enforcement is useful only when its position in the request path is clear. An agent can possess a valid identity and still propose an operation outside the intended task. The assessment should identify which component sees that operation, which rule it evaluates and whether it can refuse the request before the tool executes it.

This is narrower than the agent identity comparison. Identity records, token issuance and action checks serve different purposes. A comparison that labels all three runtime security can conceal an important gap: a decision made when a connection starts may not consider the parameters of a later tool call.

NIST's design tenets help explain the distinction. The reported detail calls for runtime verification of each tool invocation and separation between reasoning and access policy. These are design goals, not a statement that every product describing an agent identity already implements the same enforcement path.

Read the Okta gateway claim with its release status

Okta's 22 September 2026 release describes Agent Gateway as sitting between the agent and tool call. It says the gateway enforces policy and logs every interaction at runtime. Its NIST checklist adds that the gateway injects the credential at call time so the agent never holds it. That is a specific architectural claim worth testing.

The same release places Agent Gateway general availability in Q3 and expanded Kill Switch capabilities at the runtime layer in Q4. It does not state whether those are fiscal or calendar quarters. Do not infer a completion date from the quarter labels or turn a planned capability into a currently available one without a later supporting source.

The Jamf and Okta account dated 9 November says the gateway kill switch lands in Q4. It describes revoking active tokens and ending sessions in flight when an agent is deactivated at the gateway. It separately says console deactivation already blocks new sessions. These are different stopping boundaries, not interchangeable descriptions of the same effect.

Describe the Microsoft boundary without filling gaps

The Microsoft Entra documentation cited on this page describes Conditional Access applying at sign-in and token issuance. It documents dedicated agent identities, owners and sponsors. Those are useful identity controls. The cited material does not describe per-tool-call checks, so it should not be presented as evidence of the same gateway execution path Okta describes.

Microsoft's 19 October announcement adds important context. It discusses additional identity patterns, including local agents and previously unknown agents discovered through security sensors. Project Perception shares signals, context and a unified registry with Microsoft Agent 365, alongside new Playbooks and Guardian Agents. Those statements broaden the security picture without specifying an equivalent tool-call enforcement point.

Microsoft gives no availability dates for those October additions. The careful conclusion is limited: Microsoft's sourced material supports identity and token decision boundaries, while the exact action-level boundary remains not stated here. That is not a claim that Microsoft cannot enforce an action or that every Microsoft deployment behaves identically.

Trace a request through the pilot

  • Draw the actual path from agent to authorization service to tool. Mark where credentials appear and which component can inspect operation parameters.
  • Demonstrate a permitted operation and a refused operation using the same valid agent identity. Capture the deciding rule and the tool's observed result.
  • Change the resource or requested operation after token issuance. Check whether a fresh decision occurs before execution, rather than assuming the earlier decision covers it.
  • Deactivate the agent during a bounded test. Distinguish blocking a new session from ending an existing session and stopping an already requested operation.

Keep a separate entry for each tool and credential path in the evaluation record. If the available documentation does not identify the component that refuses a call, ask for that detail before declaring coverage. Record the answer as not stated until evidence resolves it. This keeps a proposed architecture separate from a verified deployment. These are proposed acceptance tests, not additional vendor capabilities. Keep the tool configuration and deployment scope in the results so that a demonstration on one path does not become a general claim about all applications. If a call bypasses the proposed enforcement component, record that path as a separate coverage question.

Ask for both positive and negative evidence. A successful tool request shows that a connection works. A denied request with a corresponding policy record shows more about enforcement. Neither result alone establishes how every tool, alternate credential or unattended workflow will behave. Scope the conclusion to the tested path.

Evaluate an alongside governance layer on the same terms

NewCore is an agent governance layer that runs alongside the identity provider a company already has. Its announcements describe runtime policy on every tool call, task-scoped credentials, human approval for high-risk actions and immediate revocation. These claims address the execution boundary directly. They should face the same request-path and failure tests as identity-provider features.

The NewCore profile provides the wider sourced product context. Keep employee sign-in and agent action policy as separate responsibilities during a pilot. Adding a governance component does not by itself establish coverage of every call. Ask how the chosen workflow reaches it and what happens when it cannot return a decision.

The governance alongside your identity provider guide explains that responsibility split. The runtime evaluation should finish with a map of covered calls, enforcement points and stop behavior. That is more actionable than selecting a product based only on whether its announcement uses the word gateway or describes an agent registry.

Frequently asked questions

Is token issuance a tool-call check?

Not necessarily. Token issuance can establish an authorized connection and permitted scope. A tool-call check examines a later operation at its execution boundary. The Microsoft sources used here describe the first boundary without establishing the second for every call.

Is Okta Agent Gateway generally available?

The September release says general availability is planned for Q3, without stating fiscal or calendar quarter. We found no later confirmation that this milestone occurred. Keep the architectural description separate from the planned release status.

How should a runtime pilot end?

Record which requests passed through the enforcement point, which rules decided them and what happened after deactivation. Include a refused operation and a changed request. This produces a bounded evidence record rather than an unsupported conclusion about all agent traffic.