Chatting with Agents
Chat is where you give direction, ask questions, and approve or push back. Every agent has one persistent thread with you, visible from its detail panel.
Send a message
Click an agent node or a sidebar entry. The chat panel opens on the right. Type a message, press ⌘ Enter to send. You get Markdown, fenced code blocks, file attachments (drag in from Finder), and an auto-save draft per agent.
What you can say
- Tasks: "Migrate the auth module to passkeys."
- Clarifications: "You said X — did you mean Y? Why?"
- Reviews: "Look at the diff I just pasted. What would you change?"
- Approvals: "Yes, push it." "No, hold — add a test first."
Managing conversations
Right-click any message for pin, copy, or split thread. Pins stay at the top of the conversation so important context does not scroll away. Full text search runs across every map with ⌘ K. Export a thread to Markdown from the chat header menu — useful for PR descriptions.
While the agent is working
The chat stays live while the agent thinks. You see the current sticky-note status, a streaming token counter, and — for long-running actions — an inline progress indicator. Press Esc to pause, or type a follow-up message; the agent will incorporate it on the next turn.
Good habits
- Decompose. Break a five-step task into three two-step tasks. Shorter turns mean faster feedback.
- Write acceptance criteria. "Done when: all unit tests pass and the README has a new section." Saves a round trip.
- Respect the memory budget. Paste small diffs instead of whole files. Let the agent ask for more if it needs more.
- Close the loop. When a task is done, say so — the Manager uses that signal to mark dependent tasks ready.