Safety / Tool access

How to Choose AI Agent Tools and Permissions

Tools turn instructions into effects. The right toolset is the smallest set that can produce the required output, show its evidence, and stop before a consequential action needs a person.

01

Begin from the output contract

List the files, facts, draft, report, or artifact the workflow must produce. Work backward to the minimum reads, writes, network requests, and integrations required.

A general browser, terminal, inbox, and cloud drive may be convenient, but broad access creates unnecessary blast radius. Prefer a focused read API or scoped folder when it can do the job.

  • Required output
  • Minimum source access
  • Minimum write destination
  • No unrelated integration
02

Separate read, draft, and execute

Read access gathers evidence. Draft access creates a reversible artifact. Execute access changes an external system or reaches another person. Treat those as separate permission tiers.

A workflow that prepares an email does not automatically need send permission. A deployment review does not automatically need production credentials.

  • Read-only first
  • Local draft second
  • External action last
  • Fresh approval at execution
03

Scope credentials and destinations

Use separate test and production credentials, narrow API scopes, short-lived tokens where available, and allowlisted hosts. Store secrets outside the conversation and skill package.

Validate filenames, paths, recipients, repositories, sites, and accounts immediately before action. A correct command against the wrong target is still a failure.

  • Test account first
  • Least-privilege scopes
  • Allowlisted hosts
  • Target confirmation
04

Add tools through evidence

Test one tool at a time with a harmless fixture. Record expected inputs, output shape, failure behavior, data retention, cost, and rollback.

Remove tools that are not used. Review access after platform updates, workflow changes, team changes, and incidents rather than letting permission accumulate forever.

  • Focused fixture
  • Failure and timeout test
  • Data and cost disclosure
  • Quarterly access review
Put it to work

Build the controlled foundation first.

Agent Ready turns these principles into four working documents before it asks you to automate anything.

Get Agent Ready Free
Continue the field manual

Related guides