Guides

Authorize AI agents to specific processes

A practical pattern for giving an AI agent access to exactly the processes you choose, using Pulse and group-based authorization.

The strategy article Prepare your business for AI-driven solutions makes the case for putting an authorization and execution layer between AI and your systems. This guide is the practical follow-up: how to actually scope an agent to a chosen set of processes with Pulse.

The model in one picture #

You never grant an agent raw access. You give it an identity, put that identity in a group, and let the group decide which processes it can run.

  • A key authenticates the agent. One key, one agent.
  • The key is bound to an agent identity (a machine user that cannot sign in).
  • The agent identity belongs to groups.
  • A group grants the right to run specific processes.

The agent can run exactly the processes its groups allow, and nothing else. This is the same authorization people get, so you only learn one model.

Why scope through groups #

The list of tools an agent can see is built from the same permission check that guards running a process directly. So the tool list and the actual run are governed by one gate. A process outside the agent's groups is invisible and cannot be triggered, even with a guessed name, and the error never reveals that another group's process exists.

That lets you hand different agents different keys, put each key's identity in a different group, and confine each agent to its own set of operations, while the credentials behind those operations stay shared and centrally managed.

A practical rollout #

  1. List the processes you want an agent to be able to run.
  2. Create a group per agent role (for example "support-agent" or "reporting-agent") and grant it only those processes.
  3. Create the agent's key and assign it to that group.
  4. Keep each process deterministic: the agent decides what to run, the workflow decides how.
  5. When a process acts for a real person, read the caller identity that Pulse carries into the run, and confirm the audit trail records who asked.

Where next #