Part 2 — For the Developer

Ch. 12 — What Agents Are Allowed to Do — Permission Models and Action Boundaries

Designing the action layer — what agents can call, write, send, and execute, and how to enforce those limits.

Knowing who the agent is solves half the problem. The other half is what it can do. These are separate concerns. Conflating them is a design error that shows up in almost every first-generation agent system.

Identity governs what data an agent can see. The action layer governs what the agent can do with it. An agent authenticated as a service account with broad read access can still be prohibited from sending messages, executing code, or writing records. Designing these constraints separately gives you two independent control surfaces — and two independent audit trails.

Most agent frameworks make this easy to ignore. A single line of configuration attaches any tool to any agent. That convenience is a trap. A document review agent that also holds a send_email tool is an agent that can exfiltrate data under the right prompt conditions.

Platform Agentic

Compliance, governance, and accountability for teams building agentic AI systems.

Access the book — sign in with Google·LinkedIn