Pilot — Supervised agentic coding. You steer.
The agent proposes plans, edits, and diffs. You review every change before it lands. Pilot is for the work where you want help — not autonomy.
How Pilot works
Supervised agentic coding. You steer.
Pilot is the mode for the work that still needs your judgement. The agent reads the codebase, builds a plan, and proposes changes — but every diff lands in front of you for review. You can accept, refine, or reject inline; you can edit the agent's plan mid-flight; you can take the wheel at any moment.
Under the hood, Pilot uses the same multi-agent runtime, the same MCP integration, and the same skills system as Autopilot. The difference is the control loop: Pilot pauses for human review at every meaningful step. Autopilot does not.
Most engineers spend most of their time in Pilot. It is faster than typing, slower than handing it the wheel — and it produces code you actually understand, because you saw every line before it landed.
When to use Pilot
Reach for Pilot when:
The situations where Pilot earns its keep.
Refactors that touch business-critical code
Pull requests where you need to defend every line in review
Working in a codebase you are still learning
Prototyping where the design is still in flux
Pair-programming with the model on hard problems
Pilot in numbers
Supervised, but fast.
Pilot proposes diffs; you accept or reject inline. Acceptance rate climbs as the agent learns project conventions.
Diff review over six weeks
Accepted vs rejected edits
Edits per week, by language
Top five — past 30 days
Get Started
Pilot ships with Parser.
Both modes share one desktop, one runtime, one set of skills. Switch without leaving the window.