Agent Continuation Atlas

Task-completion continuation

copilot --autopilot

GitHub Copilot CLI Autopilot

Give Copilot a bounded task once. It keeps initiating the next agent step locally until it declares the task complete, hits a limit, encounters a blocker, or you interrupt it.

What Autopilot controls

In a normal interactive session, Copilot eventually returns control to you for the next instruction. Autopilot changes that outer conversation flow: it can start successive model interactions on its own while working through the same task.

It does not automatically grant every tool permission. Continuation and permission are separate controls, even though Autopilot works most smoothly when the required permissions are granted up front.

01 / AssignDescribe one bounded task
02 / ActCopilot uses tools and makes progress
03 / DecideMore work needed?
04 / EndComplete, blocked, limited, or interrupted

How to start it

Inside an interactive Copilot CLI session, use Shift + Tab to cycle through standard, plan, and Autopilot modes. For a direct invocation, pass the task with the Autopilot flag.

Interactive workflow
# 1. Start Copilot CLI
copilot

# 2. Use Shift+Tab to enter plan mode and refine the plan
# 3. Choose “Accept plan and build on autopilot”
Bounded programmatic run with full permissions
copilot --autopilot --yolo \
  --max-autopilot-continues 10 \
  -p "Implement the approved plan and run the relevant checks"

Treat --yolo as a separate risk decision

It is an alias for full permissions. It can allow file changes, deletions, tool use, paths, and URLs without another approval prompt. Use a trusted repository and review the final diff.

How it stops

Autopilot continues until one of four events occurs:

  1. Copilot determines the task is complete.
  2. A problem prevents meaningful progress.
  3. You press Ctrl + C.
  4. The configured continuation limit is reached.

By default, completion returns the CLI to standard interactive mode. The stayInAutopilot setting keeps the mode selected for your next prompt; it does not make the completed task continue indefinitely.

Autonomy is not permission

Autopilot
Allows successive agent continuations until the task ends.
--allow-all
Approves tool, path, and URL access but leaves the normal interactive conversation flow in place.
--no-ask-user
Suppresses clarifying questions but does not provide Autopilot’s repeated model continuations.
--max-autopilot-continues
Caps how many continuation messages Autopilot may initiate.

Where it fits

Good fits

  • Implementing a detailed, reviewed plan
  • Writing or repairing test coverage
  • Bounded refactors with clear verification
  • Fixing a known CI failure locally

Poor fits

  • Vague product discovery
  • Work requiring frequent judgment calls
  • Untrusted repositories with broad permissions
  • Periodic monitoring over hours or days

How it differs

Autopilot is task-driven

Unlike Claude /loop, the next iteration is not scheduled by elapsed time. Unlike Codex /goal, the product description centers on autonomous completion of the assigned task rather than a durable, user-defined stopping condition evaluated across turns.

First-party sources

Verified against GitHub documentation