The first agent

An agent is the command-line program running inside a panel, with the whole terminal working.

Installing the agent program

If Claude Code isn't on the machine, the new agent's panel shows Install now. The installation is local and asks for no administrator password.

While the panel says Preparing the agent…, the app is finding out where the program lives. It takes a few seconds.

Other agent programs run in the same canvas: Codex, Grok and opencode. Each one is enabled in Settings → Providers. See several providers.

Creating

• ⌘N. It works even with the cursor inside another agent, and the new panel is born with the cursor in it.
• ⌘⇧N creates a Codex agent, and the key only does something with Codex enabled in Settings.
• ⌘K, then type agent.
• The new-agent button in the canvas bar.
• With voice set up, ⌥C and "open an agent".

⌘T opens a terminal in the same place, with no agent on top of it. See the terminal panel.

The name

Each agent is born with a name of its own. That is how voice, ⌘K and the other agents address it, and it is the name that appears in the cast after the agent is dismissed.

The first task

Type in the panel the way you would type in your own terminal. "Explain what this project does" is a short, checkable request: it reads the files, answers, and you see the panel change state along the way.

The panel's state

Each panel's indicator has five states: idle, working, waiting for you, crashed and exited.

Crashed is a turn ended by an error; idle is a turn completed. Automatic recovery of a crashed agent is configured in agent settings.

⌥F flips every panel at once and shows each agent's summary: the mission, what it did and how long ago. See knowing what they're doing.

When it doesn't open

The panel says what is missing: the agent program not installed, or the app still locating it. For everything else, see when something doesn't work.