Today, we're introducing Workflows in Bobsled.

A Workflow is a saved execution graph you configure once — including steps, dependencies, inputs, conditions, and outputs — and then invoke in chat, on a schedule, or as part of a Skill. Workflows run on Bobsled's infrastructure, so the same sequence produces the same result every time.

Data teams already use Data Skills in Bobsled to reuse trusted prompts and business rules, but until now, the agent rebuilt the execution graph itself on every turn — choosing which queries to run, in what order, and how to format the result. That's the right behavior when a question needs judgment, and the wrong behavior when the answer needs to look the same every Monday morning. Workflows make the execution graph a primitive, so the steps stay fixed, the inputs are validated, and the outputs are reproducible.

How it works

Build with the workflow agent

Tell Bobsled what you want to automate. The Workflow agent drafts an execution graph — picking the right built-in tools, wiring step dependencies, and templating inputs so the graph parameterizes cleanly.

Workflows have access to a set of built-in steps: Run SQL against the current Data Product, Execute Python for in-graph transforms, Fetch URL, Web Search, LLM Call for structured model output, and Write File for downloadable artifacts. Workflows can also call any deployed Library Tool, so they reach the same external APIs your Skills already use.

Tune in the workflow editor

The Workflow editor opens to a graph view of the draft. Add steps from the toolbox, reorder dependencies, edit each node's arguments, and use template expressions like {{inputs.year}} or {{query-revenue.output}} to wire data between steps. Independent steps run in parallel; conditions skip steps that shouldn't run.

Test and deploy

Run the draft against sample inputs from the Test tab and inspect every step's result before promoting anything. When the graph behaves the way you want, deploy it — that promotes the active draft so the next invocation uses it. The Runs tab keeps a full history of every execution: metadata, step status, input summaries, snapshots, and pointers to any files the workflow generated.

Once deployed and enabled for a Data Product, users invoke the workflow in chat with a slash command:Workflows can also be triggered from a Data App tile, a Skill, or an external event.

What teams are building

Workflows are flexible enough to cover most recurring data tasks. A few patterns have emerged from early users:

Data quality and monitoring

• Freshness sweeps: every morning, run a parallel set of Run SQL checks against critical tables, summarize the failures with an LLM Call, and use Write File to deliver a Markdown report to the data-eng channel.
• Schema drift: a weekly Run SQL plus Execute Python diff against the prior week's column inventory, with a CSV artifact of anything that changed.

Stakeholder reporting

• Monthly revenue brief: parallel Run SQL steps pull top movers, at-risk pipeline, and customer churn signals; a Web Search step adds market context; an LLM Call composes the narrative; Write File produces a branded PDF.
• Exec readouts: a single deployed workflow generates the same governed report for every business unit by passing the BU as an input, so the format never drifts between teams.

Advanced analytics

• Audience builder: users describe an audience in natural language, the workflow translates it into a deterministic SQL segment, runs it, and writes the resulting member list to a CSV ready to push into your activation tool.
• Enrichment pipelines: a Run SQL step returns the working set, a Library Tool calls an enrichment API per row, Execute Python aggregates the results, and the final dataset is written to a file artifact — without ever putting raw rows in the agent's context window.

Getting started

Workflows are available today in Bobsled. Admins can create one from Library → New → Workflow, refine the graph with the Workflow agent, test it from the editor, and deploy when it's ready. Once deployed, enable the workflow per Data Product from Behavior → Workflows so users can invoke its slash command in chat.