“Stop when the implementation is done.”
Use Autopilot when the task is bounded, the plan is detailed, and GitHub Copilot can decide when its work is complete.
See the Autopilot workflow →A field guide to agent persistence
Same ambition.
Different clocks.
Autopilot, loop, and goal all reduce the need to keep prompting an agent. The important difference is what makes each one continue—and what makes it stop.
Three continuation patterns
Start with the sentence that best describes the work. Each pattern links to a deeper explanation, invocation examples, safety considerations, and failure modes.
Autopilot
GitHub Copilot CLI
“Take this well-defined task and keep taking the next step until you believe it is complete.”
Explore AutopilotLoop
Claude Code
“Wake up on a cadence and run this prompt again while the session stays open.”
Explore LoopGoal
Codex
“Keep making verified progress across turns until this explicit condition is satisfied.”
Explore GoalThe shared mental model
All three wrap an ordinary agent turn in an outer continuation mechanism. The outer mechanism is where they diverge.
Compare the control models
The deciding question is not “which agent is smartest?” It is “what event should cause the next turn?”
| Dimension | GitHub Autopilot | Claude Loop | Codex Goal |
|---|---|---|---|
| Primary trigger | One initial task | A time interval or self-paced wake-up | One durable objective |
| Why it continues | The task still needs another agent continuation | The next scheduled iteration becomes due | The stopping condition has not been met |
| Normal stop | The agent declares the task complete | You cancel it, it expires, or the work is done | Codex is confident the goal condition is satisfied |
| Time model | Continuous multi-step run | Repeated, time-based runs | Continuous progress across turns |
| Local requirement | Runs in the local CLI session | Machine and session must remain available | Runs as long-running Codex work in the active environment |
| Best fit | Detailed implementation plans and bounded coding tasks | Polling deploys, PRs, builds, or recurring maintenance | Migrations, refactors, experiments, and measurable targets |
Choose by stopping rule
Write down the event that should end the work. That usually reveals the right mechanism faster than comparing feature names.
“Check again in five minutes.”
Use Claude /loop when elapsed time—not unfinished reasoning—is what should trigger another check.
“Stop when this testable condition is true.”
Use Codex /goal when success is explicit and the agent should keep iterating and validating across turns.
“Keep an eye on this for weeks.”
Use a durable scheduler or cloud automation. An open local session is the wrong reliability boundary for long-lived monitoring.
Understand session-scoped loops →
Verified against first-party documentation