Agents & models
Choose the agent and model in Chat, set reasoning effort, run autonomously, and approve tool calls.
Before you send a message, choose the agent you’re talking to and the model it runs on. You can change either in the middle of a conversation, and the change applies from your next message.
The agent decides the system prompt, tools, and skills. Default agent is a general-purpose assistant that’s always available. Other agents come from the capabilities installed in the project’s sandbox, or, before the sandbox exists, the ones you selected under Configure Sandbox.
If the agent you picked fails to load, your message fails with an error rather than going to a different agent. Pick another agent, or see When something breaks.
Model and reasoning effort
Section titled “Model and reasoning effort”You can choose from the models available to your account. To add more, select Manage models from the model list. See Chat models.
For models that support extended reasoning, you can also set the reasoning effort. Chat remembers your choice for each model.
Interactive or autonomous
Section titled “Interactive or autonomous”By default, the agent stops to ask you before running tools that need approval.
Run /auto to let the agent work without stopping, up to a step limit (100 by default, or pass a number: /auto 250). Run /interactive to switch back. The setting stays with the session until you change it.
Approving tool calls
Section titled “Approving tool calls”When a tool call needs your approval, the agent waits for you. Review the call’s arguments, then:
- Approve to run that one call. The next call to the same tool asks again.
- Reject to block it. The agent sees the refusal and decides what to do next.
If the turn ends before you answer, the call never runs.
See Guard policy for how the approval rules are set.