A global default connector covers every workspace unless you say otherwise. Most people never need to think about this setting again once it's picked.
A single workspace can override that default from its own Settings, if it genuinely needs a different underlying agent than the rest of your projects — nothing else about that workspace changes.
There is a second, narrower dial beside it: the Script Model, which is what the questions inside an Agent script are asked through. It exists so the small, constant judgment calls in a loop — is this in our lane, is this worth writing up — can go to something lean, most usefully a model running on your own machine, while the workspace itself keeps building through Claude, Codex, Gemini, or Grok. Left alone, those questions simply follow the workspace's own connector.
Each of those pickers can have a Temperature beside it: Low, Medium or High. Despite the name, it isn't the sampling temperature a model generates with — it picks which model in the connector's family does the work. On Claude, Low is Haiku, Medium is Sonnet and High is Opus; on Gemini it is Flash-Lite, Flash and Pro; on Grok it is Grok Build, Grok 4.3 and Grok 4.7; on Codex it moves through the GPT line the same way. Low suits the constant small judgments, High the hard building. Left on Default, the connector runs whatever its CLI would pick on its own, unless you've set a default temperature for that connector in Settings → Connectors, where the model behind each tier can be changed too.
Connectors that run one model only — a local model, a custom endpoint, or one of the scaffolded connectors — don't show the Temperature picker at all, and switching a workspace to one of them clears whatever temperature it had. A script's own questions can ask for a temperature per call as well, which wins over anything set here.
Switching a workspace's connector doesn't touch its history — past conversations stay exactly as they were. A thread you started on one connector carries on under the new one: Frontier hands it the conversation so far, so your next message picks up where the last reply left off, whichever agent wrote it.