Skip to main content
Use Agent Chat to work with a selected agent. The left side organizes agents and sessions; the conversation shows requests, replies, activity, and linked outputs. The workspace panel lets you inspect material beside the conversation.

Start and send

  1. Select an agent.
  2. Open an existing session or choose New Session.
  3. For a new session, choose its project binding and private mode before sending.
  4. Type the request, attach source material, and send.
  5. Read the agent’s response and open relevant work items or files.
Existing sessions keep their project binding. Changing the project filter in the session list changes what you see; it does not rebind an existing conversation. See Sessions and privacy.

Add context and attachments

Upload files, paste an image, or use Context to find a project, work item, document, agent, workspace member, web URL, goal, or available skill reference. Selected context appears as chips before you send; remove an incorrect chip before submitting. Context references help identify the source. They do not grant access to an inaccessible item or authorize execution. Use the actual task ID or file link when precision matters. Open a file link or work item card in a reply to review it in the workspace panel. Use Open full page or Go to Task when you need the complete detail.

Model, thinking, and speed

Model & thinking shows the effective model and where the setting comes from: the workspace, agent, or session. Choose from the models available to your workspace. Thinking levels vary by model. For models that support it, Fast mode can reduce response time and use more provider allowance. Unsupported models use Normal speed. A saved change applies to the next turn. A response already running keeps the settings with which it started. If verification fails, read the error instead of assuming the new choice is active. On smaller screens, open More composer controls. Provider connections and defaults are covered in Model providers.

While an agent responds

Activity entries can show planning, file work, tests, browser work, and other steps. Expand relevant activity or work summaries when you need context. Use Stop to interrupt the current response. When Queue is offered, queue a follow-up to send after the current response; review queued messages and attachments before they run. An interrupted turn does not automatically complete its task. A model’s activity summary is useful progress information. Verify the linked output and Task Detail before accepting tracked work as Done.

Clipboard history and navigation

Clipboard history keeps the latest supported text items copied or cut in the Dashboard for the current chat. Select an item to reuse it. It is scoped to that conversation and is not a view of your operating system’s entire clipboard. Where offered, use message navigation and bookmarks to return to important parts of a long conversation. Session search finds conversations; context search finds workspace entities.

Read-only sessions

A session your role can read may still be read-only. AgentDesk displays this state and disables sending or message changes. Open an owned permitted session when you need to continue work.