A scheduled run used to inherit whatever you last used by hand: the room the Mac was in, and the mode on the Permission chip. A change made for one afternoon's discussion silently changed every scheduled run after it. Each schedule can now say for itself who leads and, on the Mac, in which mode.
Who leads
Where to find it. Mac: Settings → The room → Schedule → Add one, the lead menu. Web: Settings → Schedule → Add one, Lead. iPhone: New schedule → Who leads.
How to use it.
- Choose a chair: Claude, ChatGPT, Gemini or Grok, named as your room names them. As the room keeps today's behaviour: whoever leads the room when the schedule fires.
- Optionally type a model id. Blank means the chair's usual model. The id is for this schedule only; your room is untouched.
- The row says leads: Gemini, on gemini-3.6-flash, on every app.
On the Mac the chair is switched on for the run even if the room has it off; the rest of the room is as the Mac is showing it. On our servers the run's own room is your seats as they are, with that chair put in the lead.
When to use it. A nightly digest on the cheap seat; the auto-fix task on the strong one; a weekly review led by a different vendor than the one that wrote the code.
In which mode
Mac only, the mode menu on the Add card: Plan, Manual, Accept edits, Smart, Auto or Bypass. As set is whatever the Permission chip says when it fires. The mode applies to that run alone and is never saved. Your team's mode floor still holds: a schedule set below the floor fails with the floor's reason rather than running looser.
The web and the iPhone show a Mac row's mode beside its lead, read-only, since the Mac runs it.
When not to use it. A model id the chair cannot run (a Claude id on the ChatGPT chair) refuses the way any wrong model does; check the room's model menu for the ids a chair accepts.
Further reading. LLM06:2025 Excessive Agency — OWASP GenAI Security Project, 2025.