Autopilot — Fully autonomous development. You watch.
Hand the agent a goal. It plans, codes, runs the terminal, calls tools, and ships — live, in front of you. Interrupt at any step. Switch back to Pilot whenever you want the wheel.
How Autopilot works
Fully autonomous development. You watch.
Autopilot is the mode for the work you want done, not done with you. You hand the agent a goal — "add OAuth login," "migrate this to Postgres," "fix the failing tests" — and it plans the work, generates tasks, runs the terminal, calls MCP tools, and executes step by step.
Everything is visible. The plan is on screen. The tool calls stream live. The diffs land one at a time. You can interrupt at any step, ask the agent to revise the plan, or switch back to Pilot the moment you want to take over.
Autopilot is not a black box that disappears for ten minutes and returns a PR. It is a transparent runtime: you watch the work, you trust it because you can see it, and you intervene only when you want to.
When to use Autopilot
Reach for Autopilot when:
The situations where Autopilot earns its keep.
Background tasks while you focus on something else
Repetitive scaffolding (boilerplate, migrations, fixtures)
Spike branches and throwaway prototypes
Test-writing for code you already wrote
Long-running multi-step refactors with clear acceptance criteria
Autopilot in numbers
Long-running. Multi-step. Hands-off.
Autopilot spawns sub-agents, runs them in parallel, and reports structured results — without touching your main context window.
Tool calls per task by complexity
Average across recent runs
Sub-agents spawned & tokens consumed
Six-week span
Get Started
Autopilot ships with Parser.
Both modes share one desktop, one runtime, one set of skills. Switch without leaving the window.