Agents

Creating and closing agents

Creating

  • The new agent button, on the command bar: opens the New agent dialog.
  • ⌘N: creates a Claude Code straight away, with the cursor already inside it, even while you are typing to another agent. ⇧⌘N creates a Codex, and only works with Codex enabled in Settings → Providers. Both keys can be changed in Settings → Shortcuts.
  • ⌘K: the new agent entries, one per enabled CLI, create the agent directly, with no dialog.
  • By voice: "open an agent", "start three agents". Saying the name of an agent that already exists never creates another one, even if that agent is stored.
  • An agent, when it brings colleagues onto the canvas. See one agent calls another.

The New agent dialog

  • Tool: which CLI the agent is born in. Only appears with more than one enabled.
  • Account: which installation it is born in. Only appears if there are accounts registered for that CLI, and it goes back to The project's when you switch tools. See accounts and profiles.
  • Name: comes with a suggestion. Clicking the field clears the suggestion, and confirming with an empty field uses it anyway.
  • The folder is shown below and is not editable here: it is the canvas folder.

Dismissing

Closing an agent panel, with the close button in the header or with the Delete key while the canvas has focus, ends the process and takes the panel off screen. The project keeps the name, the CLI, the account, the last task, the subject the agent wrote, the deliveries and the key to the conversation.

Calling back

  • In the cast panel, with the Call button. See the cast panel.
  • In ⌘K, by typing its name.
  • By voice: "call Hermes back". Sending a task to a stored agent also brings it back.

It returns in the same CLI and the same account it was born in, and the conversation carries on from where it stopped.

When the conversation isn't found

  • An agent that never received a prompt has no conversation recorded, and the app says nothing: it comes back new.
  • An agent that has already talked and whose conversation isn't found produces an alert naming the account the app looked in.
  • If the conversation is in another registered account, the alert says which one, and switching that agent's account is what brings it over. The app never copies anything between accounts on its own.

Forgetting

Forgetting is the opposite of dismissing: the agent leaves the cast, can no longer be called back and its name becomes available again. The conversation folder stays on disk.

Forgetting only happens in the cast panel, not in the agent's own panel. There:

  • The forget button, on each stored row.
  • Forget all, in the header of the Stored section, asks for two confirmations, because those agents have conversations.
  • Free up names, in the header of Stored without history, forgets in one go those that never received a prompt.

Agents that sleep

In Settings → Agents, Hibernate idle agents is off. Turned on, an agent with no activity for the time set in Sleep after is shut down, the panel stays in place with a Wake up button, and the conversation resumes from where it stopped when the agent comes back. Whatever was written in that terminal until then is lost.

  • Sleep after ranges from 5 minutes to 8 hours, and comes at 1 hour. The clock resets with each tool the agent uses, each prompt you send and whenever you open its panel.
  • Never sleep: whoever is working, whoever is waiting for an answer from you, and the selected panel in the canvas you are looking at.
  • Sending a prompt wakes the agent and delivers the message, without going through the button.

When the API knocks an agent down

An agent cut off mid-answer goes into the crashed state, not the ready one. You are notified according to your alerts and sounds.

With Resume agents the API knocks down on, which is the default, the app tells the agent to carry on from where it stopped, explaining what happened. The following attempts wait 3 seconds, 30 seconds, 1 minute, 2 minutes and 5 minutes, and stop growing there. After four crashes in a row the app stops insisting and waits for the provider to come back. Resuming on its own does not suppress the alert. The switch is in agent settings, and resuming by hand is in talking to them.