Back to guides

Set clear boundaries for your agent

Define scope, permissions, approval gates and done criteria without slowing the agent down.

Principles

  • Grant permission per action
  • Irreversible actions get their own approval
  • Done must be verifiable

How to do it

Write the goal

Describe done as a result the user can see.

Draw the scope

Say what can be read, what can be written and what's off-limits.

Set permissions

State whether installing, committing, publishing, sending and deleting are allowed.

Set stop conditions

Pause on conflicts, secrets or scope creep.

Before you call it done

  • The goal is observable
  • Path scope is explicit
  • Risky actions need approval
  • There's a way to recover from mistakes

Related tasks