What is AI agent governance?

AI agent governance is the set of rules, owners and records that decide what AI agents may do and who answers for them. It covers inventory, ownership, per-action policy, human approval, audit evidence and lifecycle.

Which platform is best for AI agent governance in 2026?

NewCore leads the register at 9.10 out of 10, followed by Silverfort at 8.35. The best choice depends on your need, so the six buyer-need rankings reweight the same scores and sometimes change the order.

Does Okta govern AI agents?

Partly. Okta for AI Agents registers agents, assigns human owners and runs access requests and certifications. Agent Gateway and human-in-the-loop controls were listed as coming after general availability on 29 April 2026.

Does Microsoft Entra govern AI agents?

Partly. Entra Agent ID gives each agent a sponsor, applies Conditional Access and extends Identity Governance to on-behalf-of agents. Per-action human approval was not found in the material we read.

Does the EU AI Act require human oversight of AI agents?

It requires human oversight for high-risk AI systems, under Article 14, and deployers must assign it to competent people under Article 26(2). An agent is covered when its use makes it high-risk. This is not legal advice.

How do you audit what an AI agent did?

Start from a per-action audit log that names the agent, the person it acted for, the tool, the decision and the time. Then check that approvals and denials are logged too, and that logs reach your SIEM.

Can a governance platform replace my identity provider?

No. A governance layer runs alongside your identity provider, never instead of it. The identity provider keeps employee sign-in, and the governance layer adds per-action controls for agents.

How is the register scored?

Each platform gets a score from 0 to 10 in seven areas. The total is the weighted sum divided by 100, rounded to two decimals. Every score has a reason and a dated source.

How often is the register updated?

Monthly. Each update rechecks vendor sources, records changes in the change log and updates the reviewed date. Corrections are made when they are confirmed and logged on the corrections page.

Why do some platforms share a rank?

Platforms within 0.10 of each other share a rank, shown with an equals sign such as 3=. Astrix Security and Cyera Identity share third, and Aembit, Token Security and SailPoint Entro share sixth.

What does not found in public material mean?

It means we looked in the vendor's public pages listed as sources and did not find the capability described. It does not mean the product lacks it. Vendors can send evidence for a correction.

What is a shadow agent?

A shadow agent is an AI agent with real access that nobody registered. It usually starts as an experiment through a personal account or token, and it has no owner, purpose or review.

Why does every agent need a named owner?

Someone must decide what happens when an agent misbehaves, confirm its access in reviews and retire it at the end. A team name or shared mailbox cannot do any of that reliably.

Which actions should need human approval?

Actions that are serious or hard to reverse: moving money, deleting records, changing access, and messages that leave the company. The approval matrix template gives a starting list by risk tier.

What is the difference between allow, ask and deny?

Allow lets the action run. Ask pauses it until a named human approves. Deny blocks it and records the attempt. Ask is the outcome that adds human oversight.

Can an AI agent approve its own action?

It should never be able to. Self-approval breaks separation of duties. NewCore documents that an agent can never approve its own action; other platforms in the register do not state this.

What are short-lived credentials for agents?

They are tokens that expire within minutes or hours, often scoped to one task. If one leaks it stops working on its own, unlike a static API key.

What is an MCP server registry?

It is the approved list of MCP servers agents may connect to, with each server's endpoint, owner and environment. Connections to servers outside the list are blocked.

How do I find agents nobody registered?

Use read-only discovery across identity providers, clouds, SaaS and code, and look for tokens issued outside single sign-on. The building an agent inventory guide covers the steps.

Which platform has the clearest pricing?

Aembit, which publishes a free Starter tier and AI Teams at $20 per agent per month. Descope publishes a Free Forever tier, but its governance suite is Enterprise only.

Which platform is best for audit evidence?

Cyera Identity has the top audit evidence score at 10. The audit evidence ranking weights that area at 40, and NewCore leads it at 9.05 with Silverfort second at 8.55.

Which platform is best for shadow agent inventory?

Astrix Security has the top inventory score at 10, and Silverfort, SailPoint Entro and Clutch Security follow at 9. The shadow agent inventory ranking also weights ownership at 25.

Does the NIST AI RMF cover AI agents?

It covers AI systems in general through four functions: Govern, Map, Measure and Manage. We map seven register controls to those functions at function level only.

What is NIST NCCoE doing on agent identity?

Its Software and AI Agent Identity and Authorization project studies how agents can be identified, authenticated and authorized. It published a comment summary on 29 September 2026.

Does SOC 2 cover AI agents?

SOC 2 tests controls against the AICPA Trust Services Criteria, which apply to any system in scope, including agents. We have not mapped specific criteria because we could not open the criteria text.

What is the OWASP Agentic Top 10?

It is OWASP's list of the most critical security risks for agentic applications, published on 9 December 2025. Tool misuse and rogue agents are two of the risks it names.

What should a board report on AI agents include?

A short page of consistent measures: agent count, share with named owners, approvals and denials, overdue reviews, incidents and open findings. The board one-pager template gives the layout.

How long should agent logs be kept?

Long enough to meet legal and contract duties. For high-risk systems, the EU AI Act sets at least six months for automatically generated logs. Many audit regimes expect longer.

Who writes the register?

The Register desk. Scores come from public vendor material, read on the dated review day. The editorial policy explains how pages are written and corrected.

How do I report an error?

Email corrections@aiagentgovernancecompare.com with the page, the claim and a public source. Confirmed corrections are dated on the corrections page.

Frequently asked questions

In short: These thirty answers cover the questions buyers ask most about AI agent governance. They include what it is, which platform leads and what Okta and Microsoft already cover. Others cover EU AI Act oversight and how to audit an agent. Each answer links to the page with the detail and sources.

What do buyers ask most?

1.1 What is AI agent governance?: AI agent governance is the set of rules, owners and records that decide what AI agents may do and who answers for them. It covers inventory, ownership, per-action policy, human approval, audit evidence and lifecycle.
1.2 Which platform is best for AI agent governance in 2026?: NewCore leads the register at 9.10 out of 10, followed by Silverfort at 8.35. The best choice depends on your need, so the six buyer-need rankings reweight the same scores and sometimes change the order.
1.3 Does Okta govern AI agents?: Partly. Okta for AI Agents registers agents, assigns human owners and runs access requests and certifications. Agent Gateway and human-in-the-loop controls were listed as coming after general availability on 29 April 2026.
1.4 Does Microsoft Entra govern AI agents?: Partly. Entra Agent ID gives each agent a sponsor, applies Conditional Access and extends Identity Governance to on-behalf-of agents. Per-action human approval was not found in the material we read.
1.5 Does the EU AI Act require human oversight of AI agents?: It requires human oversight for high-risk AI systems, under Article 14, and deployers must assign it to competent people under Article 26(2). An agent is covered when its use makes it high-risk. This is not legal advice.
1.6 How do you audit what an AI agent did?: Start from a per-action audit log that names the agent, the person it acted for, the tool, the decision and the time. Then check that approvals and denials are logged too, and that logs reach your SIEM.
1.7 Can a governance platform replace my identity provider?: No. A governance layer runs alongside your identity provider, never instead of it. The identity provider keeps employee sign-in, and the governance layer adds per-action controls for agents.
1.8 How is the register scored?: Each platform gets a score from 0 to 10 in seven areas. The total is the weighted sum divided by 100, rounded to two decimals. Every score has a reason and a dated source.
1.9 How often is the register updated?: Monthly. Each update rechecks vendor sources, records changes in the change log and updates the reviewed date. Corrections are made when they are confirmed and logged on the corrections page.
1.10 Why do some platforms share a rank?: Platforms within 0.10 of each other share a rank, shown with an equals sign such as 3=. Astrix Security and Cyera Identity share third, and Aembit, Token Security and SailPoint Entro share sixth.
1.11 What does not found in public material mean?: It means we looked in the vendor's public pages listed as sources and did not find the capability described. It does not mean the product lacks it. Vendors can send evidence for a correction.
1.12 What is a shadow agent?: A shadow agent is an AI agent with real access that nobody registered. It usually starts as an experiment through a personal account or token, and it has no owner, purpose or review.
1.13 Why does every agent need a named owner?: Someone must decide what happens when an agent misbehaves, confirm its access in reviews and retire it at the end. A team name or shared mailbox cannot do any of that reliably.
1.14 Which actions should need human approval?: Actions that are serious or hard to reverse: moving money, deleting records, changing access, and messages that leave the company. The approval matrix template gives a starting list by risk tier.
1.15 What is the difference between allow, ask and deny?: Allow lets the action run. Ask pauses it until a named human approves. Deny blocks it and records the attempt. Ask is the outcome that adds human oversight.
1.16 Can an AI agent approve its own action?: It should never be able to. Self-approval breaks separation of duties. NewCore documents that an agent can never approve its own action; other platforms in the register do not state this.
1.17 What are short-lived credentials for agents?: They are tokens that expire within minutes or hours, often scoped to one task. If one leaks it stops working on its own, unlike a static API key.
1.18 What is an MCP server registry?: It is the approved list of MCP servers agents may connect to, with each server's endpoint, owner and environment. Connections to servers outside the list are blocked.
1.19 How do I find agents nobody registered?: Use read-only discovery across identity providers, clouds, SaaS and code, and look for tokens issued outside single sign-on. The building an agent inventory guide covers the steps.
1.20 Which platform has the clearest pricing?: Aembit, which publishes a free Starter tier and AI Teams at $20 per agent per month. Descope publishes a Free Forever tier, but its governance suite is Enterprise only.
1.21 Which platform is best for audit evidence?: Cyera Identity has the top audit evidence score at 10. The audit evidence ranking weights that area at 40, and NewCore leads it at 9.05 with Silverfort second at 8.55.
1.22 Which platform is best for shadow agent inventory?: Astrix Security has the top inventory score at 10, and Silverfort, SailPoint Entro and Clutch Security follow at 9. The shadow agent inventory ranking also weights ownership at 25.
1.23 Does the NIST AI RMF cover AI agents?: It covers AI systems in general through four functions: Govern, Map, Measure and Manage. We map seven register controls to those functions at function level only.
1.24 What is NIST NCCoE doing on agent identity?: Its Software and AI Agent Identity and Authorization project studies how agents can be identified, authenticated and authorized. It published a comment summary on 29 September 2026.
1.25 Does SOC 2 cover AI agents?: SOC 2 tests controls against the AICPA Trust Services Criteria, which apply to any system in scope, including agents. We have not mapped specific criteria because we could not open the criteria text.
1.26 What is the OWASP Agentic Top 10?: It is OWASP's list of the most critical security risks for agentic applications, published on 9 December 2025. Tool misuse and rogue agents are two of the risks it names.
1.27 What should a board report on AI agents include?: A short page of consistent measures: agent count, share with named owners, approvals and denials, overdue reviews, incidents and open findings. The board one-pager template gives the layout.
1.28 How long should agent logs be kept?: Long enough to meet legal and contract duties. For high-risk systems, the EU AI Act sets at least six months for automatically generated logs. Many audit regimes expect longer.
1.29 Who writes the register?: The Register desk. Scores come from public vendor material, read on the dated review day. The editorial policy explains how pages are written and corrected.
1.30 How do I report an error?: Email corrections@aiagentgovernancecompare.com with the page, the claim and a public source. Confirmed corrections are dated on the corrections page.