Overview
How might we make agent autonomy a visible, selectable policy instead of a hidden autopilot?
When to use
- Coding agents, workflow automation, and tools where the same run can draft, edit, or ship.
- Sessions that mix reversible local work with irreversible external actions.
- Products that need a shared vocabulary for “how independently is this agent allowed to act?”
- Any surface where silent autonomy would surprise people after the side effect lands.
When to skip
- Single low-stakes draft tools where confirm always and execute never is enough.
- Fully attended interactive editors where every keystroke is already human-driven.
- Regulated flows that legally require confirm for every side effect, expose confirm only, not execute.
States
Design the mode policy and the step outcomes together, not only a three-way toggle.
- 01
Mode set
Suggest, confirm, or execute is chosen before work starts. The active mode is always visible.
- 02
Suggesting
The agent proposes a plan or next action. Nothing consequential runs until a human starts it.
- 03
Awaiting confirm
A step is staged with inspectable scope. Execution waits for an explicit approve or skip.
- 04
Executing
Prior permission covers the run. Steps progress with visible status instead of silent background work.
- 05
Paused
Mode change, error, or a riskier step stops the run and re-discloses what would happen next.
- 06
Complete
The workflow finished or stopped. Outcome is separate from which mode authorized it.
Key UX elements
The parts that must be present for autonomy modes to feel trustworthy.
Modes
Name the three policies in plain language.
Suggest only, ask first, and run through. Labels must map to real behavior, not marketing.
Disclosure
Say what the current mode will do.
One sentence under the control: propose only, pause per step, or continue without stopping.
Plan
Show the steps before side effects.
List tools and targets so people can see blast radius before any mode runs them.
Gate
Confirm pauses on the active step.
Approve and skip are first-class. The staged step stays inspectable until a decision.
Progress
Make execute visible while it runs.
Step status replaces a silent autopilot. People should see what finished and what is next.
Default
Start safer than the maximum.
Default to confirm for new users and irreversible actions. Offer execute after scope is understood.
Rules
A global Autopilot that silently includes irreversible execute actions.
Mode labels without showing what will run (tools, targets, blast radius).
Defaulting newcomers to execute on first session.
Switching modes mid-run without pausing or re-disclosing side effects.
Evidence
| Product | Implementation |
|---|---|
| Cursor Agent | Plan and ask before apply vs freer edit modes with reviewable diffs. |
| GitHub Copilot Workspace | Proposal and plan stages before merging agent-authored changes. |
| Claude Code | Permission prompts for tool use vs trusted local edit sessions. |
| Devin | Staged plans and human gates before long autonomous coding runs. |