Skip to main content
Chat is how you talk to any agent. Dispatch is how agents run and talk to each other. This page covers both.

​
Composing and sending

  • Type in the input box and press Enter to send. Shift+Enter adds a line break.
  • The input auto-grows as you type, up to 50,000 characters.
  • Hover your own message to Copy it or Edit it (editing loads the text back into the input for you to revise and resend).
  • Type / to open the slash-command menu with the agent’s installed skills. Picking one inserts the command into your message. See Skills.

​
While an agent is working

When the agent is streaming, the input switches to steer mode and three controls matter:
  • Steer: send a message mid-run to redirect the agent without stopping it. The placeholder reads “Send a steer message while the agent is working…”.
  • Stop: cancel the current run for this agent (does not pause the team).
  • Answer Now: interrupt the run and ask the agent to respond with what it has so far.
A live status card shows elapsed time, a phase label (“Thinking”, “Reading your codebase”, “Planning approach”, “Writing code”, and so on), and the last few tool actions (Read, Edit, Bash, Grep, WebSearch, and others), so you always know what the agent is doing.

​
Attachments

Three ways to attach files:
  1. Drag and drop onto the input (the border highlights and the placeholder changes to “Drop files to attach…”).
  2. Paste an image or screenshot from the clipboard.
  3. The Attach file button (plus icon) opens a file picker; multiple files at once are fine.
Supported types include images (PNG, JPG), PDF, DOCX, and XLSX. Images show as thumbnails (click in chat to open a full-screen lightbox); other files show as removable chips. In a team workspace, attached files also collect in the bottom tray’s Uploads tab so the whole team can find them later.

​
Rich output

Agent responses render as full markdown:
  • Code blocks with syntax highlighting, a language badge, and a Copy button.
  • Tables that can become charts: when a table has numeric data, a View as chart toggle appears. Pick Bar, Line, Area, Pie, Scatter, or Radar, choose axes and series, stack or flip bars, expand to fullscreen, and Save as PNG or Save as JPEG.
  • Status badges for lines starting with ✅ ❌ ⚠ 📋.
  • File links: file names render as clickable badges that open a read-only preview with line numbers.

​
Approvals in chat

Depending on the agent’s permission mode, a Tool Request modal can appear before an agent acts, showing the tool, the exact command or file, and Allow / Deny & Stop buttons. Destructive actions are flagged: “Destructive action — review carefully.” In team chat you may also see interruption cards when Atlas wants to message another agent or open new work, with Allow & Send / Reject buttons. See Tasks and Working with Atlas.

​
How dispatch works

A dispatch is one agent run. It starts when:
  • You send the agent a message.
  • Atlas assigns it a task.
  • A teammate sends it a message: if the recipient is idle it wakes immediately; if it is busy, the message queues and the agent re-dispatches when it finishes.
During a run the agent sees its Soul, recent conversation, relevant team messages, and team context, then works in the working directory and reports back. Its tool activity streams live into the chat and the Activity tray.

​
Agent-to-agent messages

Coordination traffic is kept out of your chat. Read it in the Dispatch tab: the whole team’s coordination feed. In Atlas chat, the Show updates from agents checkbox controls whether short “Update from [agent]” digests appear inline as teammates report progress. The setting is remembered per team.

​
If something fails

Failed sends show an error below the input and restore your message and attachments so you can retry. An agent that errors out goes red (“Blocked”); open its chat to see the reason, then message it again or cancel its task. More: Troubleshooting workspaces.