Choose the model a workflow runs on, and pay with your own key

Every agent loop resolved the same three things as platform-wide constants: which model, how hard it thinks, and whose key pays. So a workflow that needs deep reasoning and one that needs to be cheap were the same workflow, and nothing recorded what had actually run.

You can now pick a model and a reasoning effort per workflow, separately for running it and for the interview that authors it. Every model on offer supports computer use, because a run drives a real browser and the alternative is finding that out on its first screenshot.

Settings takes your own OpenAI key. Once it is saved, it routes everything of yours: runs, the agent watching a codified run, the interview and the reviser. Your prompts sit under your own contract and the inference lands on your own bill.

Model and effort are pinned to a run when it starts, so a run that waits two days on an approval resumes as the same agent that suspended, even if you have changed your mind since. Your key is the opposite: it is read fresh every time, so work that queued up behind a broken key drains on the fixed one. A key that dies halfway through a run lets that run finish on ours rather than dying in the middle.