Skip to main content

​
DeepAgents Workspace

DeepAgents can organize multi-step work inside a conversation workspace. You can follow that work through Tasks, Files, Preview, and Memories in the Chat right drawer. There is no separate Deep Mode toggle. Workspace content appears when the selected Agent and its tools produce it.

​
Tasks

The Tasks panel displays the live to-do state reported by the Agent. It distinguishes pending, active, and completed items so you can follow a multi-step run. The Chat task list is read-only. Guide the work by messaging the Agent or by approving or rejecting a requested action. Scheduled automation is managed separately on the Tasks page.

​
Files

The Files panel combines files arriving during the current run with files available for the open thread.
  • Browse folders in the file tree.
  • Open text and code in the inline viewer.
  • View supported images, audio, and other media.
  • Copy file content or a media path.
  • Export Markdown as PDF, optionally with the company logo.
The first available file opens automatically on desktop. On smaller screens, select a file and use the back control to return to the tree. Files shown here belong to the conversation workspace. Availability after a later visit depends on the thread and workspace services. Copy or export important results instead of treating the workspace as guaranteed long-term storage.

​
Preview

The Preview panel is for supported generated artifacts:
  • website workspaces can start or wake a live preview
  • presentations can render their entry file directly
  • ordinary file collections show that no live preview is available
You can refresh the preview or open its URL in another tab. A preview may need time to start, may be recreated after inactivity, or may be unavailable when runtime capacity is limited. The panel does not guarantee that arbitrary code or every file type can run.

​
Memories

The Memories panel groups information for the selected Agent and user:
  • Core profile contains structured profile information.
  • Long-term contains durable facts and preferences.
  • Conversations contains per-thread summaries; in Chat, this layer is filtered to the open thread.
You can add a long-term memory and manage existing entries. Memory creation is selective. A message, file, or task is not automatically a durable memory unless the Agent’s memory behavior stores it.

​
Approvals and control

When a tool requires human review, Chat displays an Action required card with Approve and Reject. The Auto-approve actions setting can bypass these prompts for the current run configuration. Approvals govern requested actions; they are not a switch that enables DeepAgents or guarantees autonomous completion. Keep auto-approval off when you need to review external or sensitive operations. Scheduled tasks have a separate workspace drawer with user-created to-dos and files plus files generated by the latest run. This is different from the live task state shown in Chat. See Tasks for that workflow.