AI UX PlaygroundNewsletterJoin 2K+ AI designers and PMs on Substack. New teardowns, patterns, and prompts as they drop.

Agents

Per-Action Autonomy

Set independence per capability, not one global switch. Auto-draft email but ask before send. Auto-commit on a branch but never to main.

Interactive demo

Tool permissions

Choose how each tool runs. Changes apply to this project only.

edit_file

Edit files in the workspace

run_terminal

Run terminal commands

git_push

Push to remote

delete_file

Delete files

Overview

How might we design per-action autonomy so people can trust and act on AI output?

When to use

  • Essential for AI coding agents, workflow automation, and productivity assistants where different actions carry different risk profiles and users need calibrated autonomy per capability, not per app.

When to skip

  • Single-capability agents where per-action settings duplicate a global toggle.
  • Audiences who will never open advanced settings; offer sensible defaults first.
  • When the platform cannot enforce different policies per tool call.

Rules

  • One “Autopilot” switch that silently enables irreversible actions.

  • Per-action settings buried with no summary of current policy.

  • Defaults that auto-approve send/delete on first connect.

  • Policies that drift after model or connector updates without notice.

Evidence

ProductImplementation
Email agentsAuto-draft known threads; approve send to new recipients.
Cursor / coding agentsFreer edits in working trees; gated pushes to protected branches.
ChatGPT agent modeDifferentiated confirms for browsing vs messaging vs purchases.
Enterprise IT botsAuto-read; approve write to production systems.

FAQ

What is per-action autonomy in AI agents?

Per-action autonomy sets how independently each capability may run, so read, draft, and send can have different approval rules inside the same agent.

Why not a single autonomy slider?

Risk is not uniform. A global slider either blocks useful automation or over-authorizes dangerous actions. Per-action policies match real trust gradients.

How should defaults be set?

Default to ask-to-approve for irreversible or external actions, and allow freer autonomy for reversible local drafts. Make the policy matrix visible and editable.