Skip to main content
A run is one autonomous task: a prompt executed as a headless session in a fresh VM on a workspace. Dispatching a run returns immediately with a run ID; the run finishes with a summary and any files the agent staged in ~/outputs/.

​
Start a run

  • Runs on different workspaces execute in parallel; runs on the same session key serialize (one live session per workspace).
  • --session selects the workspace: same key means the conversation continues — the transcript is resumed from the volume. No key means a fresh anonymous workspace (72h TTL).

​
Check on a run

status shows run, agent, status, model, effort, PR, summary, and error. list supports filtering with --status (e.g. queued, running, completed, failed). You can also watch or steer a live run from its session timeline in the dashboard (run page → Open Session Timeline).

​
Run statuses

​
Continue a run

Dispatches a follow-up run on the finished run’s workspace — full conversation and files intact. This is the only way to continue an anonymous workspace, and the standard way to iterate (“fix the review feedback on the PR you opened”). It accepts the same --model, --effort, --size, and --timeout overrides.

​
Outputs

When the final turn ends, the run completes and the VM is gone. Deliverables:
  • Summary — shown in run status and on the run page.
  • Files — anything the agent staged in ~/outputs/ is snapshotted onto the run at finalize. Every VM boots with ~/outputs/ created and the convention written into the agent’s platform instructions, so agents stage user-facing files there without being told.
  • Messages — anything sent via send_message lands in the agent’s Inbox and your configured channels.