Talking to the agents

Typing

Click the panel and type. It is the CLI's own terminal, so its slash commands and line editing work in there.

Dictating

Dictation delivers the text to the selected agent, without going through the assistant. With no panel selected there is no destination, and the dictation bar says so. See dictation.

By voice

"Tell Hermes to run the tests", "ask Atlas whether that compiled".

• The text goes as you said it, with no paraphrase.
• A stored agent is brought back with its whole conversation, instead of being born again.
• Two similar names stop the delivery: the app asks which of the two before sending anything.
• An app verb does not become a task: "close Perseu" closes his panel instead of sending the word into his terminal.

Answering what they ask

In Settings → Conversation:

• Answer agents by voice: after saying aloud that an agent is waiting, the app opens the microphone for you to answer without touching anything. Turned off, it keeps alerting you and leaves the notification on screen longer, without opening the microphone.
• Response window: how many seconds the microphone stays open after the alert, from 0 to 15, at 6 by default. At 0 the listening is off. The control requires voice enabled and the switch above turned on, and the listening happens on your Mac, with no cost per second.

What is read aloud, and how to answer by picking an option, is in knowing what they're doing.

Spotlighting a panel

"Focus on Hermes" puts his panel large on the left and rearranges the others on the right. It works for any panel, and the ones with no name of their own are called by type: "focus on the browser". "Unfocus" returns everyone to the grid.

Telling a crashed agent to carry on

"Tell Apolo to carry on" is a request of its own, different from typing "continue" in the terminal: the message sent explains what happened and asks the agent to check what was left half done before redoing anything.

• It respects the wait between attempts: asked during the wait, the app says in how many seconds it will send.
• "When Claude is back" puts the agent in the queue and only resumes when the provider leaves the instability.
• "Leave it alone" cancels the automatic recovery for that agent.

See creating and closing.

The history of what it delivered

In Settings → Agents, The agent's delivery history comes on: each subject the agent completes becomes a line on the back of its panel when it moves to the next subject. Turned off, the back shows only what it is doing now.

Commands that never end

In Settings → Agents, Dev servers in their own panel comes off. Turned on, long commands such as a dev server or a file watcher are opened in a visible terminal panel, where you see the log and can stop them. The classification of long commands gets it wrong sometimes.

Telling an agent to hand off the work

With agents isolated in worktrees, the workstreams sidebar carries a hand-off button per workstream. It merges nothing: it sends the message to the agent inside that copy of the project.

In Settings → Interface, in the Workstreams sidebar section:

• How to hand off a finished workstream: what the message asks for. Infer from each destination is the default and looks at how work usually lands on that branch; Direct merge, Pull request and Just notify apply to every destination in every project. The choice made in the sidebar's own menu overrides it, and only for that destination.
• Edit the message: the text sent to the agent. It has three tokens, which become the branch name, the destination and the verb of the chosen mode. Without the verb token, the message doesn't say what to do.

See worktrees and isolation.

Talking to an agent is not the same as an agent talking to another

An agent talks to whoever it brought onto the canvas itself, and returns the result to whoever brought it. See one agent calls another.