Approve or Auto, and the other four settings that decide what runs on your machine
The engine's settings panel is short on purpose. Here is each setting, what it defaults to, and why the defaults are conservative.
Most orchestration tools have a settings page you need a doc for. The engine's has six controls, and we spent longer removing settings than adding them. This is what is left and why.
Running work
Start items: Auto or Ask me. The same switch as the menu bar panel, stored here. Ask me is the default. The engine will arm items and tell you they are ready; it will not start one.
At once. How many sessions the engine runs in parallel. Two numbers apply, and the panel shows both: the seat's limit, set by whoever administers the workspace under People, and the engine's, set here. A run on this seat alone uses the seat's number; a route that spans more seats uses the engine's. The default is 1, and most people go to 3 or 4 once they have watched a few waves land.
Runs on. Which seat this engine runs as: your Claude Code, your Codex, your Cursor. One engine, one seat; a second seat is a second engine, and the board shows them apart.
Opens sessions in. tmux, Orca, a terminal you name, or headless. tmux is the default because the watch handle is tmux attach and the menu bar app can open it for you. Headless is for a machine you never look at.
Models
Three rows, because the engine does three different jobs with a model and they do not need the same one.
- Builds. The model that runs a node. This is the one the band was computed against, so changing it changes what "in band" means; the record notes it.
- Decides. The model that writes the decision when a node parks: the question, the two or three options, what each one means for the band. Defaults to the same as builds.
- Drafts. The model that writes the route itself from a request, and the field book the nodes carry. Defaults to the same as builds, but the usual shape is a stronger drafter over lighter builders: one careful pass that splits the work well, then cheaper sessions that do what it said.
Each row has an effort setting beside it, and "Agent default" leaves it to the harness.
Where the settings live
Your profile is stored on sleds, not on the machine, so a second Mac running your engine behaves the same. A flag on sleds engine wins over the profile for that one run, which is how you try a model for an afternoon without changing your defaults.
Why conservative
Every default above is the one that does less: ask before starting, one at a time, the harness's own effort. An engine that runs agents on a person's machine should earn the next setting, not assume it. The panel makes flipping any of them a second's work once it has.