Agent settings

In Settings → Agents. What is here applies to every project. What belongs to a single project (the account its agents are born in, the model, the permissions, the default browser) lives in the canvas settings. See project settings.

• The agent's delivery history: on, each completed subject becomes a line on the back of the panel when the agent moves to the next subject. Off, the back shows only what it is doing now.
• Resume agents the API knocks down: on, the app tells the agent to carry on from where it stopped after an API crash, waiting longer with each new attempt. The crash alert happens whether it is on or off. See creating and closing.
• Hibernate idle agents: off by default. On, an agent with no activity for the time below is shut down, the panel keeps a wake button and the conversation resumes when the agent comes back.
  • Sleep after: appears inside the control above, and only with it on. It ranges from 5 minutes to 8 hours, and comes at 1 hour. The clock resets with each tool used, each prompt sent and whenever you open the agent's panel. Never sleep whoever is working, whoever is waiting for your answer and the selected panel in the canvas in focus.
• Claude installations: registers .claude folders, each with its own login, servers and conversations. Add installation… asks for the folder. Removing one from the list deletes nothing from disk, and the app asks what to do with the agents that depended on it. See accounts and profiles.
• Claude Code version: The latest keeps Claude Code updated on its own when the app opens, before any agent is born. The menu also lists the versions present on the machine, and pinning one turns off automatic updating. The change applies to agents created from then on.
• Pull request badge: shows each agent's PR on the panel, in the footer and in Git, with the state of continuous integration. Off, the app stops asking GitHub about your PRs. See the Git panel.
• Dev servers in their own panel: off by default. On, long commands such as a dev server or a file watcher are opened in a visible terminal panel, instead of running inside the agent. See talking to them.
• Isolation by worktree: on, each agent creates its own copy of the project before writing and commits on its own branch. The text of the protocol the agents receive is editable right below, and the editor only appears with the switch on. See worktrees and isolation.
• Subagent panels: Show subagents on the canvas gives a panel to each task an agent delegates, up to three at a time. Off, subagents keep working without showing up. See one agent calls another.
• Conversations between projects: how many messages one authorization of yours allows between an agent here and an agent in another project. The options are 2, 4, 6, 10 and 20 messages, and the default is 4. Once used up, they need a new authorization.

In the Interface tab

• Glow on the panel when done: the panel border glows when the agent delivers, even without being selected.
• Highlight the area tab when done: the area tab flashes when an agent finishes, or starts waiting for you, in an area you are not looking at.
• Completion glow color: the color of the two above. The Default button only appears after you change it, and puts it back.
• Group the agents into tabs, in the Workstreams sidebar section: in an area with more than one agent, they share a single panel, with the chosen one on show and the others in tabs. It applies to areas never configured by hand; each area's context menu overrides it. See work areas.
• Hide the area tabs while it's docked, in the same section: the tabs at the top disappear and the workstreams sidebar takes over listing the areas, switching on click, creating, renaming and flashing. The control only works with the sidebar docked.