An AI agent that actually works

Tools tied to the active workspace

The runtime can expose document, filesystem, table, web, browser-context, image, artifact, and shell tools. The exact list changes with the model, provider, pane, and task instead of being one universal catalog.

Visible plans and review points

Plan events appear as pending, active, and completed steps. Supported table writes can require an explicit Apply action, while shell commands can pause for Allow once, a remembered prefix, or Deny.

Context you can inspect

Attach the current document, another workspace item, a file, or up to eight read-only local folders. The context tray makes it clear what may be sent to the selected model provider.

Stop and verify

Stop aborts the active stream, changed-file cards expose affected paths, and artifacts stay linked to the run. Completed actions are not automatically rolled back, so source-level verification still matters.

What a workspace Agent does

A workspace Agent uses selected project context and available tools to work on documents, records, files, and other supported objects. Its useful result is the work you can inspect, not merely a fluent answer. Describe both the intended output and the constraints that matter for the task.

Follow a reviewable task

Use the timeline to understand what happened and inspect the actual destination afterward. Different surfaces have different review controls, and stopping a stream does not undo actions that already completed.

1. Open the intended workspace and select relevant context.
2. Describe a bounded outcome and identify what must remain unchanged.
3. Inspect tool activity and respond to the applicable review or approval controls.
4. Verify the final document, record, image, or artifact in its destination.

• Agent basics
• Whiteboard proposal review