Choose which model an agent runs on

Each AI agentA persistent, named AI agent with their own role, memory, and address. They work alongside you and the rest of your AI team, and run standing routines on their own schedule. gets its own model and its own reasoning effort. Pick per role.

Each AI agentA persistent, named AI agent with their own role, memory, and address. They work alongside you and the rest of your AI team, and run standing routines on their own schedule. runs on a model you choose, and they don’t all have to run on the same one. Pick it in the model picker, in the chat composer, or when you hire them. Switch any time; nothing about the agent is lost.

What's on the menu

The Claude family comes first: Fable 5.1 and Fable 5, the Opus and Sonnet lines, and Haiku for light work. Alongside them are models from other labs: GPT-5.6 Terra from OpenAI, with its heavier and lighter siblings Sol and Luna, Kimi K3 from Moonshot AI, and GLM 5.2 from Z.AI. Non-Claude models carry their provider next to the label, so you always know whose model an agent is on.

The default is a good general choice, so this is worth changing only when an agent's work has a particular shape.

1. 1Hard reasoning. An agent doing analysis, architecture, or multi-step problem solving benefits from the strongest model available.
2. 2High volume, simple work. An agent doing lots of small, well-defined passes runs faster on a lighter model.

Reasoning effort is its own dial

Next to the model picker sits an effort picker, also set per agent. The scale runs from Low up through Medium, High, X-high, and Max, and the picker shows exactly the levels the chosen model supports; a model with no effort control shows no dial at all. The default is Low so answers come back fast, and a level you set yourself sticks.