AI agent Beta Optional feature

The AI agent takes natural language instructions and autonomously performs data analysis and notebook creation and editing. It can reference the catalog and notebook content while running SQL, checking results, and creating charts.

The model, permissions, and fixed context it uses are managed through an AI profile. For preparing context that's loaded automatically, see Agent context.

If you're using the legacy AI Assistant, see AI Assistant.

INFO

This feature is in beta. The specification might change without notice.

TIP

This is an optional feature. Depending on your contract, it might not be available.

Key features

• Sending data to the LLM: How data is handled for training purposes, and what information might be sent to the LLM.
• Enabling it: Enable the AI agent under AI feature settings in workspace settings.
• Display modes: Side panel / full page.
• Canvas: View, share, and download pages created by the AI agent.
• Operation modes: Ask (analysis) / Agent (including creation and editing).
• Specifying context: Use @ to specify tables, notebooks, saved queries, and more.
• Agent context: The .agent folder or page, and context from an AI profile.
• Message queue: Queue, edit, or send a message ahead of turn while a run is in progress.
• Plan & Todos: Review the Plan and Todos for the current session.
• Conversation history summary: Summarize a long conversation history.
• User settings: Auto-approving SQL, auto-loading context, language, and more.
• AI profile: Manage the model, permissions, and fixed context used.
• Agent usage: Check session and token usage.
• Agent connectors: Integrations with external tools (MCP / API).

Sending data to the LLM

Only external LLM services that contractually guarantee your input data won't be used for training are used. If you want to use your own API key, you can configure this under Custom in AI profile.

As the AI agent analyzes and processes data, it might send table data, SQL execution results, and similar information to the LLM service. Check your organization's AI usage policy before using this feature.

Enabling it

Enabling AI features requires Workspace Owner permission.

1. Open global nav > Workspace settings > AI feature settings.
2. Turn on AI features overall using the switch at the top of the screen.
3. Turn on AI Agent under Available AI features.

You can also individually enable the following features on the same screen.

Feature	Description
AI Agent	Performs data analysis, SQL execution, chart creation, and more from natural language
AI code completion	Provides inline SQL completion within a SQL block
AI name suggestion	Automatically suggests names for SQL and charts

Opening it

You can open it in any of the following ways.

1. Open global nav > AI Agent (full page).
2. Open the AI Agent icon at the bottom right of the screen (side panel).
3. Open the side panel with the AI Assistant shortcut (⌘+L / Ctrl+L by default; see Keyboard shortcuts).
4. Select Mention in the AI agent from a SQL block in a notebook.

If it isn't enabled, AI Agent is disabled is shown, and you can go to the enablement screen from Manage settings.

Display modes

Display	How to open	Description
Side panel	AI Agent at the bottom right of the screen	Makes it easy to use the notebook you're viewing as context
Full page	Global nav > AI Agent	Lets you view chat content in a larger area. You can also get here from Session info > Open in full page mode in the side panel

Canvas

The AI agent can present analysis results and other content as an HTML Canvas. Each Canvas receives its own URL when it's created and remains available after its original session ends.

You can open a Canvas preview from the chat. In the preview, you can select Open in a new tab from the menu to view the displayed version on its dedicated page. The owner can switch versions, and switching versions also updates the page URL to that version.

A URL without a version opens the latest version. If you can only view the shared version, such as a viewer or someone using a public link, that version is shown and the version is removed from the URL. The owner can change sharing settings from Share.

In the AI agent's full page, the preview appears beside the chat. From the Canvas in chat, the preview, and the dedicated URL page, you can select Download HTML or Download PDF from Download to save the content. The PDF uses A4 portrait, and people viewing a public link can download it without signing in.

Get the dedicated URL from Share > Copy link. Opening that URL shows a page with only the Canvas. If the Canvas is shared with Anyone with the link, people can view it without signing in. Opening a restricted Canvas shows a sign-in path that returns to the same URL.

If you are signed in and can use the workspace, the logo at the top left takes you to the workspace home. In a workspace where the AI agent is available, Workspace Viewers and higher see the AI agent launcher. Workspace Guests do not see it.

When you open the AI agent from a Canvas page, it uses the displayed Canvas as context. Owners can ask the agent to read or edit the Canvas. Viewers can ask it to read the Canvas but not edit it.

The owner can also change sharing settings from this page and switch versions. If the AI agent session open on this page creates a new version, the page switches to the latest version and drops the version from the URL. It doesn't switch when the agent isn't open or a different session is open.

If you need to set up multi-factor authentication, the Canvas page shows the setup path regardless of whether the Canvas is public. If a temporary problem prevents the workspace from loading, you can retry from the same URL.

Sharing settings

The owner can change these settings under Share.

Item	Description
People with access	Grants view access to users, groups, or all workspace members.
General access	Choose Only people with access or Anyone with the link. If you select Anyone with the link, people can use the same URL without signing in. You can't select it if workspace policy prevents public link creation.
Shared version	Selects the version that viewers see. While Anyone with the link is selected, a fixed version is required, so you can't select Latest.

Editing a Canvas doesn't automatically move a fixed shared version forward. To share newer content, change Shared version to the newer version. When people other than you can view the Canvas, the version switcher labels the version they see as Shared. If Latest is selected, the current latest version gets the label.

Operation modes

You can switch between Ask and Agent at the bottom left of the chat input field.

Mode	Description
Ask	Performs data analysis, including running SQL. Doesn't edit notebooks or the catalog
Agent	Performs every operation, including creating and editing notebooks. Before editing, an AI edit version is created automatically, so you can revert it

Users without edit permission on the notebook can't select Agent and are limited to Ask.

Specifying context

In the chat input field, use @ to specify what you want the AI agent to reference.

Target	Description
Table	A table in the catalog
Notebook	A notebook in the workspace
Saved Query	A saved query
SQL Block	A SQL block in the notebook you currently have open (when candidates exist)
Chart	A chart in the notebook you currently have open (when candidates exist)

Chat screen actions

You can perform the following actions from the header.

Action	Description
New chat	Starts a new session
Chat history	Lists past sessions and lets you switch between them
Session info	Check the start time, Total tokens, and more. From the side panel, you can also select Open in full page mode
Close	Closes the side panel (in side panel mode)

In the chat input field, you can change the following settings in addition to the operation mode.

• The connection to prefer (default is Auto, letting the agent choose)
• Settings (user settings)
• AI Profile
• Language

Message queue

If you send a message while Thinking... or Running is shown, a waiting list appears above the input field, under the heading Queued messages. Once the current run finishes, queued messages are processed in order starting from the first one. While there's text in the input field, the send button reads Add to queue.

Action	Description
Send now	Interrupts the run in progress and sends that message immediately
Edit	Edits a queued message. While editing, use Save / Cancel
Delete	Removes the message from the queue

Plan & Todos

You can review the work plan (Plan) and todo list (Todos) that the AI agent created in the session. Open it when the Plan & Todos button is shown near the chat input field.

What's shown

Item	Description
Plan	The Plan saved by the AI agent. Switch the display format with Rendered / Raw. If none is saved, No Plan has been saved yet. is shown.
Todos	Todos for the current session. You can check each item's content and status (pending / in_progress / completed). If none are active, No todos are currently active. is shown.

Conversation history summary

Enter / in the input field and select Summarize conversation history to summarize the conversation history in the current session.

You can continue adding new messages to the queue while the summary runs. Select Stop generating response to stop the summary.

User settings

You can change these from Settings > User Preferences in the chat input field.

Item	Description
Enter to newline	When on, Enter inserts a line break and a modifier key+Enter sends. When off, Enter sends
Show SQL details	Shows the SQL body within an Agent message by default
Auto-approve Tool Execution	Run SQL (per connection; for BigQuery, you can also set Size Limit (BigQuery) in GB), MCP Tool Execution, Connector HTTP Request
Auto-load Agent Context	Personal / Teamspace / Notebook page (all on by default)

Permissions

The AI agent's actions are carried out within the user's permission scope. The notebooks it can explore and the connections it can use are also limited to what the user can access.

You can configure additional restrictions under Tool permissions in the AI profile (notebooks, connections, connectors, web search / fetch).

If you want to confirm every SQL execution, turn off Run SQL under Auto-approve Tool Execution in user settings.