Building with the assistant

Letting the assistant edit the page you are on, build pipelines, and deploy with your confirmation.

The assistant works on the page you are looking at. Open Assistant from the right edge of the window, and on the pages below it can read what is on screen and change it for you. On any other page it still answers questions, queries your data and builds pipelines, but it leaves the page itself alone.

Pages it can edit

Page	What the assistant can change	When the change is kept
Query	The SQL in your tabs. It can also open a new tab or switch to another one.	Right away. Nothing runs until you press Execute.
Home	Node positions and sizes, zones, the search and tag filter, Snap to grid and the selected node. Connections, datasets and pipelines stay as they are.	Right away, in this browser. See Data landscape.
Pipeline editor	Sources, including Trigger on update and the run mapping, SQL stages, the destination, variables, and the run values used for Preview.	When you save.
Automation editor	Name, Instructions, What should the report include?, starting queries, Use previous runs, the model and reasoning effort, and the schedule and timezone.	When you choose Save automation.

For the pipeline editor, open it on its own page with Open full page, or beside a conversation in AI Chat.

What you see after an edit

The fields, tabs or stages the assistant changed flash briefly, and its reply ends with a Draft updated row. Dismiss puts the page back the way it was before that turn. On the pipeline editor the row also has Save, which does the same as the editor's own Save button.

In the pipeline and automation editors, nothing is stored until you save, so you can read the draft first. On Query and Home the change applies at once, and Dismiss is how you undo it.

Dismiss covers the most recent turn that changed the page. The row goes away once you save or move to another page.

Build a pipeline in conversation

Ask for the result you want

Describe the output and name its inputs. Type @ to mention a dataset by name. For example: "From @orders and @customers, build a pipeline that writes daily revenue per country to a Gold dataset called daily_revenue_by_country."

Let it explore and test

The assistant looks up the datasets, samples their columns and runs its SQL as queries before it builds anything. These queries are read-only.

It creates the pipeline

The assistant creates the pipeline with its sources, stages and destination as the Working version. Nothing is deployed and nothing is written. In AI Chat the new pipeline opens in the editor beside the conversation.

Review and preview

Read each stage and press Preview on it. Preview runs your current draft and writes nothing. See Builder tour.

Refine, then save

Ask for changes while the editor is open. The assistant now edits your draft in the editor, and you keep it with Save.

Whether the pipeline is open decides where an edit lands. With the pipeline open in the editor, the assistant changes your unsaved draft and leaves the stored pipeline alone, so your own unsaved edits are never overwritten. When it is not open, the assistant changes the stored Working version directly. That change takes effect at once and chat cannot undo it, though the Deployed version keeps running unchanged until someone deploys again.

The assistant's own previews run against the saved Working version. After it edits a draft you have open, preview the changed stages in the editor yourself.

Deploy from chat

Ask the assistant to deploy, and the turn pauses on a card that names the pipeline, the Working version and the destination's write mode. Choose Allow to deploy or Deny to stop. Deploying is the only action the assistant asks about first.

Allowing a deploy promotes the Working version to Deployed and starts a run that writes the destination dataset. Beetl runs the same checks as the Deploy button; see What Deploy checks.

Three cases stop the deploy:

The pipeline is open with unsaved changes. The assistant asks you to Save or Dismiss the draft first.
The card is more than 10 minutes old.
The pipeline changed between the card appearing and your Allow. Nothing deploys, and the assistant tells you.

MCP clients use the same deploy, but a client with a write key deploys without asking anyone, so treat such a key like a deploy credential. See The agent trust boundary.

Automations from chat

In chat, the assistant works on automations through the automation editor. Open New automation, or an existing automation, and ask it to fill in the check, for example "check every morning that no order in @orders has a negative amount, and list the offending order IDs". It writes the instructions, starting queries and schedule into the form. Choose Test automation to try the draft, then Save automation to keep it.

Running, pausing and deleting saved automations happens on the Automations page, or from an MCP client with a write key.

Ask for what you want

Specific requests get better results:

Name the datasets and columns you mean, with @ mentions where you can.
Ask for one change at a time, such as "add a stage that drops cancelled orders", and check each result before the next.
Ask for one stage per logical step, so you can preview every intermediate result.
Ask it to preview a stage and show you the rows before you build on them.
Ask it to explain the SQL it wrote before you save it.
State the write mode you want. A pipeline the assistant creates replaces the whole destination table on every run (Overwrite) unless you ask for Append or Merge.

Review before you save or deploy

Each stage reads the sources and earlier stages you expect, and its preview shows the rows and types you want.
The destination has the right dataset name, storage tier, write mode and partition columns. See Write modes and partitioning.
Trigger on update is on only for sources that should start runs, and the run mapping supplies every run.* value.
The deploy card names the pipeline and write mode you expect before you choose Allow.