Skip to content

Choose a runtime, model and mode

Every agent runs on one runtime, such as Claude or Codex, and each runtime offers its own choices: which model, how much it asks before acting (its mode), and sometimes how hard it thinks (effort). This guide sets them for a new agent, changes them for one that is running, and reads the meter that says how full its context is.

Before you start

Steps

For a new agent

  1. Unfold the project in the sidebar and click New session, the first row under it (or press ⌘N).
  2. Click the runtime above the prompt, on the right, and choose one. A runtime that cannot be used says why under its name, such as Needs signing in.
  3. Under the prompt, the runtime's own choices appear as capsules. Their names and values are the runtime's, so they differ from one to the next. Usually:
  4. the mode, on the left, such as Ask for approval or Full access for Codex, or a plan or read-only mode;
  5. on the right, the model, and effort where the runtime has it.

Click one and choose. A runtime that offers no choices, such as Cursor, shows none.

Cursor and Grok also have no permission-mode capsule under the prompt: set Default or Always-approve for each in Settings ▸ Agent Runtimes. Cursor's Agent, Plan and Ask stay as they are; they are not that permission setting.

Beside the mode is the command sandbox, such as Sandbox: runtime controlled or Sandbox off. It is the app's, not the runtime's list, and it is separate from the mode: the mode says when the agent asks you, the sandbox says what its commands can reach. Choose Use runtime default to follow Settings ▸ Agent Runtimes, or On or Off for this agent only. For Codex, Off is Full access and turns its approval prompts off too; the mode capsule changes with it. A runtime the app cannot change shows its state only; hover over it for why. 4. Type the prompt and send it. The agent starts with what you chose.

On iPhone and iPad, the start form has Runtime, Sandbox, Mode, Model and Effort rows instead.

For an agent that is running

  1. Open its conversation.
  2. Change the mode, model or sandbox in the same capsules under the prompt. A mode or model change applies from the agent's next step, a sandbox change from its next turn, never in the middle of a command; the runtime itself cannot be changed. To carry the work on with another runtime, start a new chat there and ask it to continue this one; see Keep going when a runtime runs out.

Read the meter

Above the prompt of a running agent, beside its name, is what it has cost so far, and a ring where the runtime says how big its context is.

  • The ring fills as the conversation fills the model's context. Hover over it for the numbers. A nearly full context is the most common reason an agent starts to forget or wander; start a new agent, or branch this one, before it fills.
  • The cost is the runtime's own figure, never an estimate. It turns red when the agent is close to a spending limit; hover for how much is left. A runtime that reports no price says so, and no limit applies to it. See Limit what agents spend.

A workflow sets the same choices in its file, with runtime:, model:, effort:, permission-mode: and options:. See Workflow triggers and actions.

If it doesn't work

  • The runtime I want is not in the menu. It is not installed on this Mac, or on the server the project is on. See Runtimes.
  • There is no ring. The runtime does not say how big its context is.
  • "Claude's sandbox could not start" (or another runtime's) over a new agent's prompt. The runtime cannot set up its sandbox on this computer or server. The prompt stays in the box; Start without sandbox sends it again with this agent's sandbox off. See Answer a question or a permission request.