Authorization was built for a different kind of actor.

Software authorization was designed around three kinds of principal:

Each is either accountable in person or predictable by construction. A role granted to them means roughly the same thing tomorrow as it does today.

AI agents are different.

• They interpret.
• They plan.
• They choose tools.
• They process untrusted information.
• They act probabilistically.
• They learn from context.
• They may operate continuously.
• They may modify their own workflows.

The same agent can be reliable at 9:00 and compromised by a document at 9:04. Its model can be swapped for a cheaper one without anyone updating its permissions. A static grant can’t see any of this.

Static IAM alone is insufficient.

Identity and access management remains necessary. But agents need a layer that decides, at the moment of action, how much of that access should be exercised without a human.

What the next layer requires.

1. 01Identity
2. 02Policy
3. 03Context
4. 04Evidence
5. 05Risk
6. 06Dynamic authority
7. 07Human escalation
8. 08Verification

Access with accountability.

Governance is the product.