AI Agent Runtime Authorization & Control Tools

When to explore this layer

Look here when a proposed agent can perform consequential actions and you need a decision at execution time. A policy written in a manifest is only useful if the execution path checks it. Draw the full path from agent to tool to target system before assessing coverage.

ILLUSTRATIVE EVALUATION · NOT A CUSTOMER RESULT

Put a real task in the demonstration.

For a fictional procurement agent, allow preparation of a supplier draft but require a separate approval for activation. Try the activation through each available tool path, alter the payload after approval, and make the authorization service unavailable. The demonstration should reveal bypasses and the documented failure behavior.

Questions to bring to the demonstration

1. Where is the decision enforced, and can another connector bypass it?
2. Is approval tied to the exact resource, action, payload and policy version?
3. Does a failed policy lookup deny, defer or allow the action?
4. How do expiry, replay prevention and emergency revocation work for pending actions?

Evidence to request

• An enforcement-path diagram
• Denied bypass and altered-payload attempts
• Decision logs identifying the policy and approval versions

Record what was demonstrated, what was only described, and what remains unknown. Preserve the product version, environment and date beside each observation.

Where this layer stops

Runtime controls evaluate actions in their coverage. They do not establish whether the company's policy is correct, resolve an unclear business owner, or guarantee that an allowed action produces a good outcome.

Connect it to the work

An explicit work unit makes the policy review concrete: prepare this artifact, read these resources, and stop before this decision. That is the upstream specification DutyGraph aims to help an advisor assemble.

Read our perspective on the demand side of agents →