Background jobs and agents
Background work continues after the assistant’s foreground reply ends. The gateway owns the job and delivers its result back into the originating thread. This is useful for a long test suite, transcription, research, or a build you do not want to hold open as a chat turn.
Delegate an agent task
Section titled “Delegate an agent task”Give a complete brief, including inputs, output, and boundaries:
Run a background agent to review the authentication code in this workspace.Write a report with file references and proposed fixes; make no code changes.Post the completed report in this thread.A background agent starts a separate engine session in the same conversation workspace. It follows the conversation’s permissions and can be used in any mode. It receives a self-contained task rather than the foreground thread’s entire context. Both Claude and Codex are supported.
The launch returns a job ID. The report arrives when the agent finishes; you do not need to keep sending “continue.” A background agent may run for up to a week, subject to its runtime and operational limits.
Delegate a shell command
Section titled “Delegate a shell command”A background shell job is appropriate when the command itself needs to outlive the turn, such as a long build. It executes in the conversation’s resolved runtime, outside the engine’s own tool approval loop.
Shell jobs require Auto, or Admin mode with a current admin author. Auto still requires an admin to approve the exact command on the durable Run it card. Admin-authored jobs in Admin mode can start directly. Prefer a confined background agent when the work benefits from ordinary engine permission handling.
Track completion and recovery
Section titled “Track completion and recovery”Ask the assistant for the job’s status or use /status in Slack to inspect running work. Keep the returned ID when discussing an individual job. A shell completion can resume the parent assistant to review output; an agent report is delivered directly.
Job records persist. A surviving detached shell process can be recovered after a restart; an interrupted background agent is reported as interrupted rather than replayed as if its tool effects were known. Persistent records do not guarantee arbitrary external commands survive service teardown.