Agent Dock
The Agent Dock puts the agent conversation in the right-hand tab strip of the Data and Design perspectives, so you can ask a question without losing sight of the sheet or canvas you are looking at.
It is the same agent as the full-screen Agent Perspective, not a lighter version of it. The two share one conversation: start a task full-screen, then tuck it into the dock while you check the result, and nothing restarts or replays.
Why dock it?
In the full-screen perspective, the conversation and the canvas trade places. The dock removes that trade-off. Ask for an analysis, watch the widget appear three tabs away, and read the answer without navigating anywhere.
Good fits for the dock:
- "What is this column actually holding?"
- "Add a total row here." — small, visible, verifiable actions
- Watching a long task progress while you keep working
- Checking a chart's data binding right after the agent built it
Reach for the full-screen perspective when you want the conversation, the trace, and the plan to be the main thing on screen.
It remembers whether you want it
The dock opens automatically when you enter the Data or Design perspective. Close it once — click the Agent tab again, or the dock's close control — and it stays closed on later entries. Open it again and the app remembers that too.
There is no setting to hunt for: the tab strip is the switch.
Sharing the strip with Properties and other panels, the dock keeps its own width, wider than the property panels because a chat needs the room. Drag the divider to make it yours; switching tabs snaps the strip to that tab's width.
What is in the panel
Everything the full-screen view has:
- Conversations — your past chats, grouped into Today, Yesterday, and Earlier. Titles are named for you once a conversation gets going; rename or delete from the conversation's menu. A spinner marks a conversation that is still working, a red dot one that is waiting on you.
- The conversation — markdown answers and a list of the steps the agent actually took.
- Trace chips — thinking and delegated agents above each reply; click to open the trace. See Thinking & Reasoning.
- Approvals — when an agent wants to write something, the request appears above the input box. See Settings → Permissions.
- The message box — agent selector, model selector, Plan/Act toggle, and a ring showing how full the conversation's context is. Hover the ring for the breakdown.
When the panel is narrow, the controls collapse into an overflow menu instead of being clipped, so the input box stays usable.
The context ring
The ring in the corner of the message box shows how full the current conversation's context window is. It is a reading, not a setting:
- It counts what the next turn has to send again — the model's most recent prompt, the answer it produced on that step, and anything a delegated agent handed back that has not been folded in yet.
- It fills as the conversation grows, turning amber past 60% of the window and red past 80%.
- Hover it for a tooltip split into two labelled sections, because they answer different questions:
| Section | Lines | Question it answers |
|---|---|---|
| Context occupancy (sent again next turn) | Final step prompt · Final step output · Subagent (this turn) · Total — used / limit and a percentage | How much of the window is taken |
| Final step billing split | Cache hit · Cache miss · Cache write | How that last call was priced |
The two do not move together. A cache hit still occupies context; a cache miss does not make the window any fuller. Only the Total line drives the ring. The Subagent (this turn) line appears only when a delegated agent contributed on that turn.
Long conversations are compacted automatically — older turns are replaced by a summary while the tool results that matter are kept. You do not need to do anything; if you want a fresh start, begin a new conversation.
Approvals still apply
Docking does not lower the bar. Any action that writes to your files asks first, exactly as it would full-screen, and delete-class operations always wait for you.
Where to go next
- Agent Perspective Overview — the agent team and how it delegates
- Plan Mode — agree on a design before it is built
- Settings → Permissions — review what you have allowed