What an agent is

An agent is an agent CLI (Claude Code, Codex, Grok or opencode) running in a real terminal inside a canvas panel.

What runs in there

• The whole program: the slash commands, the model picker, the diffs and the permission prompts are the CLI's own, and you type into them as you would in Terminal.
• The configuration already in the project: the instructions written in the folder, the MCP servers and the subagents are read by the CLI, which runs in the canvas folder.
• The binary installed on this machine: the app looks for each CLI in your PATH and shows in Settings → Providers which ones it found. See several providers.

The folder

The working directory is not chosen at creation: it comes from the canvas, and every agent created there is born in it. See choosing the project.

The name

The name is chosen at creation, with a suggestion already filled in, and it is how you address the agent by voice, in ⌘K and in the cast panel.

• While the agent exists in the project, on screen or stored, its name is reserved and no new agent can take it.
• Forgetting an agent releases the name back to the pool of available names. See creating and closing.

Permissions

In the canvas settings there is the Don't ask for permission switch, on by default. Turned off, the CLI goes back to asking before each tool, the agent stays in the waiting state until the request is answered, and you are notified. The permission mode menu only appears with it turned off. See project settings.

Its own browser

An agent can open a browser on the canvas, and that panel belongs to it: no other agent navigates, clicks or takes a screenshot there. See the agent sees the browser.

The panel and the agent are different things

Closing the panel dismisses the agent and keeps the conversation, the name, the last task and what it recorded as delivered. See creating and closing.