Models, effort and permission
Three knobs on every chat — which model, how hard it thinks, and what it may do without asking you first.
Under the message box is one row: ＋, the model, and send. The model is on
the row because it is the setting worth reading before every message. Everything
else that changes how the next message is answered is behind the ＋.
Which model
Each agent offers its own models, and the panel shows exactly the list that
agent publishes. Switch in the middle of a chat if you like; the thread
carries on.
How hard it thinks
Some agents expose an effort setting — low, medium, high, and higher. More
effort is more thinking, more time and more of your subscription's allowance
spent on one answer.
Worth turning up for a design decision or a nasty bug. Not worth it for
"rename this function".
What it may do without asking
This is the one that decides how much you have to be present.
Each agent has its own permission modes and its own names for them, and the
panel shows that agent's own list. They fall into a few familiar shapes:
- Plan — think it through and come back with a plan. Changes nothing.
- Read only / review — look, explain, answer. Changes nothing.
- Build — make the edits, and ask before anything unusual.
- Accept edits / auto — get on with it, stop only for the big things.
- Unrestricted — stop for nothing.
The ＋ tells you when you have left the loosest mode on: it wears a mark, in
the colour that means "look at this". To see which mode is actually on, open the
sheet; it is named there.
Grok Build is the exception: its permissions live in its own configuration on
the machine rather than in the panel, so its row here is effort instead.
Choosing one for being away
A mode that stops on the first question is a mode that will sit untouched until
you pick the phone up. If you want work to run to the end while you are out,
give it enough rope to get there — and give it a job where that is safe.
If you would rather be asked, the chat waits, and it waits properly. See
when an agent needs you.
The context meter
Under the message box is a small gauge: how full this agent's memory of the
chat is. Tap it for the breakdown, and for one button — compress
context — which is the agent's own summarise-and-continue.
It is only drawn where the number is real. An agent that reports what it is
using gets a meter; one that reports nothing gets none. Where an agent gives one
total rather than a breakdown, the sheet says the total and does not invent the
parts.
Watch it on a long thread. A chat that has filled up gets vaguer before
it gets any warning, and compressing early is cheaper than starting again.
Three switches that are modes, not settings
Behind the ＋, at the very top, under how this chat runs, in this order:
- Sketch first — get a clickable stub of a screen before it is built. See
seeing it before it is built.
- Hold files while editing — on by default, so no two agents on this
machine edit one file at once. See
two agents, one file.
- Put other agents to work — the crew switch, with the roles that crew may
reach hanging off it. See what a crew is.
They are at the top because the last one spends other agents' allowances while
nobody is watching.
Sketching and holding files are both the machine's rather than this chat's,
so flipping either changes every chat on that machine; only the crew
switch belongs to the chat you are in. Holding files is the one whose mark on
the ＋ appears when it is off rather than on.
A row with nothing behind it is not drawn at all, so a machine with no roles to
offer simply has one row fewer.