Orchestra: split one task across several models

Orchestra splits one task into subtasks, sends each to a model suited to its size, and has one model oversee the rest. It quotes the run in model calls before you start. It needs Personal or Business and is only partly built.

Last checked against the app on · View as Markdown

Steps

  1. Open the command palette with Ctrl+K and go to Orchestra. It is in no sidebar, and it only appears for an edition that includes the agent fleet.
  2. Read the plan panel. It states how many model calls the run may make, and that estimate is made before anything runs and makes no model call itself.
  3. Type the whole job into "The task" box, not one step. The page says it will be split into subtasks.
  4. Press "Run orchestration". The button reads "Running" while the run is with the provider, and progress per subtask appears when the app sends it.
  5. Read the result when it arrives. If the run fails, the message is the provider’s own words and anything already spent stays spent.

How to tell it worked

The page shows a result panel for your task, and the plan panel’s call count matched what ran.

If it does not work

The page says Orchestra cannot run on this install.
The message under it gives the reason. Free does not include Orchestra, and an install with no model provider configured cannot build a run.
You expected to choose a model per subtask.
You do not. The app picks a model for each subtask’s size from the provider it is set to use. Pick the provider and model in Chat first.
A banner says a previous run may still be in progress.
Starting another runs alongside it, and its result may not show on this page.