Models and budgets
The agent’s model
Section titled “The agent’s model”Each agent uses a fixed model or Automatic. With Automatic, Rook picks the model for each turn, starting from cheap, deterministic signals, and uses a stronger model only where it earns its cost.
| Tier | Used for |
|---|---|
| Fast | Small, routine turns: triage, summaries, short replies. |
| Main | Ordinary work and anything that continues an already approved action. |
| Strong | Code, large contexts and work the others failed at. |
The choice follows this order: the tier the run already settled on; code, or an engineering, research or computer agent (strong); a very large conversation (strong); continuing an approval (main, at least); an email, triage, support or summary agent with a short request (fast). If it is still ambiguous, one tiny call to the fast model names the tier; if that fails, main is used.
A run that fails on a cheaper tier because of a loop, repeated failing tool calls or unusable output moves up to the next tier once. It shows on the Usage page.
Budgets
Section titled “Budgets”An agent can have a monthly token limit (UTC month). At the limit, new runs fail before calling the model, with a clear notice. Each run also has a reply cap of 10% of the monthly limit, so a single task cannot eat the month. Whoever can configure the space can also pause every agent at once.
The server enforces the limit and the pause; no agent can read or change those settings.