Running an agent identity layer alongside Okta or Entra
In short:Keep Okta or Entra for employee sign-in. Add an agent identity layer for what sign-in does not cover. That means discovery of unregistered agents, checks on every tool call and approval of risky actions. An agent identity layer runs alongside your identity provider, never instead of it.
Do I need anything beyond my identity provider?
Maybe. Okta and Microsoft have both shipped agent features this year. Before buying more, read what they already cover.
Okta for AI Agents became generally available on 29 April 2026. It registers homegrown agents in Universal Directory and imports agents from Salesforce Agentforce, Amazon Bedrock AgentCore and ServiceNow. Short-lived tokens replace hardcoded credentials. Admins can assign human owners and run access requests and certifications.
Okta Agent SSO, built on Cross App Access, became generally available on 24 August 2026. It is included in core Okta SSO at no extra cost.
Microsoft Entra Agent ID is available to all Entra customers. Each agent has a sponsor. Conditional Access applies to agents. Identity Governance covers on-behalf-of agents. Extended features need Microsoft Agent 365 at $15 per user per month.
What is still missing?
Three gaps show up most often.
Discovery of agents nobody registered. Okta imports from named platforms. Entra creates identities for agents built on Microsoft tooling. Neither describes finding a coding assistant wired up with a personal token.
Per-call checks. An identity provider decides whether an agent gets a token. It does not usually see each tool call the agent makes afterwards. Okta listed its Agent Gateway as coming after general availability.
Human approval for risky actions. Okta listed human-in-the-loop controls as coming after general availability. For Entra Agent ID, per-action human approval was not found in the sources we read.
If those gaps matter to you, you need a layer that covers them.
What does running alongside mean?
It means your identity provider stays the source of truth for people. Employees keep signing in where they do today. Nothing about workforce SSO changes.
An agent identity layer runs alongside your identity provider, never instead of it. It reads people and groups from the directory. It adds agents, owners and policies on top. It sits in the path of agent traffic to check each call.
Several vendors say this plainly. NewCore states that your current identity platform keeps handling employee sign-in. Its first deployment path is called "Start alongside your identity platform". Aembit states it does not replace the workforce identity provider. Silverfort enforces SSO through your identity provider and reads Entra ID and Okta for discovery.
How do the pieces connect?
A typical layout has four parts.
- The identity provider signs people in and holds groups.
- The agent identity layer holds agents, owners and policy.
- A gateway sits between agents and tools and enforces policy on each call.
- An approval channel reaches named people when a call needs a decision.
When an agent makes a call, the gateway identifies the agent and the person behind it. It checks the action against policy. If the verdict is ask, the approver gets a prompt. The result goes into one audit trail.
Where does Cross App Access fit?
Cross App Access lets the identity provider decide which apps an agent may reach. It became the MCP Enterprise-Managed Authorization extension, adopted by Anthropic, Microsoft and Okta.
That handles the connection: may this agent talk to this app at all? The agent identity layer handles the action: may this agent do this thing right now? The two work together. Our page on Enterprise-Managed Authorization has the details.
What should I look for in a layer that runs alongside Okta?
Weigh runtime control highest. That is the gap Okta has said it will fill later. Our keep Okta ranking gives runtime control the largest weight for this reason.
Then check that discovery reads Okta and goes beyond it. Agents using personal tokens outside SSO will not show up in Okta at all. NewCore traces access paths outside SSO.
Finally, check that owners in the layer map to people in Universal Directory, not to a separate list.
What should I look for alongside Entra?
Weigh delegation and runtime control. Entra already gives you sponsors and Conditional Access. What it does not describe is approval on each risky action.
Check how the layer reads Entra agent objects and sponsors. Silverfort reads Entra ID through read-only APIs. Ask whether a sponsor in Entra becomes the approver in the layer or whether you set approvers separately.
Our Microsoft 365 shops ranking uses weights tuned for this case.
How do I roll it out without disruption?
Start in observe mode. Point a few agents at the gateway and log decisions without blocking. Compare what policy would have done with what actually happened.
NewCore describes writing policies in a sentence and replaying them against recent real sessions before they go live. Whatever product you pick, test policies on real traffic first.
Then turn on enforcement for one high-risk action. Deleting records, sending payments and exporting data are good first candidates. Add human approval for that action only. Expand once approvers trust the flow.
What will the identity team need to own?
The identity team usually owns the connection between the directory and the layer. They make sure groups and people sync correctly. They decide how agent owners relate to managers in the HR system.
Security engineering usually owns policies and the gateway. Application teams own their agents and answer reviews. Write this split down before the rollout, or reviews will go unanswered.
What does this cost?
Okta Agent SSO is included in core Okta SSO. Microsoft Agent 365 is $15 per user per month for extended features. Agent identity layers vary. Aembit publishes $20 per agent per month for AI Teams. NewCore sells through AWS Marketplace with custom units. Silverfort packages are quote-based.
Budget for the gateway's hosting if you run it yourself. Aembit's MCP Identity Gateway is a customer-hosted VM.
The short version
Keep your identity provider. Read what it already covers for agents. Add a layer for discovery, per-call checks and approval if those gaps matter. Roll out in observe mode first, then enforce on the riskiest actions. See the standings for how the scored platforms compare.