Skip to main content

Background runs

Work sent off to run on its own. Each run is a separate agent in its own directory, so it keeps going after you close the chat that started it.

Four ways one starts

SourceStarted by
Dispatchyou, by hand, in the Dispatch panel
Scheduleda recurring task coming due
Dutya bot's duty coming due
Coworksending a message in Cowork mode

All four end up in the same list, which is why the panel is worth reading even if you never dispatch anything by hand.

They run only while OpenCLI is open

This is a local agent, not a server. Anything that must fire while the machine is asleep belongs in the operating system's scheduler — cron, launchd, Task Scheduler.

Said plainly because the alternative is a task that silently never ran.

How many at once

Three by default. Each is a whole agent with its own model calls, so more than your machine can feed makes them all slower rather than finishing sooner. The number is a control in the panel and takes effect without a restart.

Sixteen is the ceiling, and not as a matter of taste: past it, runs stop making progress and start competing for the same weights.

Directories, and why a run may be held

A run's sandbox can write anywhere inside its working directory. That directory is therefore not a convenience — it is the whole of what the run may change.

So a run is allowed to start when its directory is one this product already knows about:

  • inside the workspace
  • inside a department's own directory
  • somewhere you have allowed by name

Anything else is held, and appears under Waiting for you with an Allow this directory button. Allowing a place releases every run waiting on it.

If your approval setting is Never ask, nothing is held. Somebody who has turned approvals off has said as much, and the product does not overrule the setting it offered.

Watching one

Output appears as the agent produces it, not when the run finishes. A run that takes ten minutes is readable for all ten.

Picking one up

A finished run can be continued as an ordinary conversation: Continue in a chat opens a new thread in the same directory, with what the run was asked and what it reported already in context.

The run itself cannot be spoken to. This is the way back in.