MotherDuck’s workshop slate treats data agents as systems that must survive broken data

MotherDuck’s Data Outpost workshop program is easy to dismiss as conference promotion. Read as an engineering artifact, though, its 11 sessions expose a useful change in how vendors are framing data agents: the interesting work starts after a model can generate SQL.

The first-party workshop guide groups the November 4 program into building agents, preparing data for them and operating them in production. Attendees can choose three sessions because the 11 labs run across three parallel time slots. That structure matters less than the repeated operational concerns inside the individual exercises.

Break the pipeline on purpose

One MotherDuck-led lab asks participants to let an agent profile files, design a schema, ingest data, write transformations and publish the result. Crucially, the exercise then breaks the data deliberately. The advertised output is not just a working stack, but prompts and guardrails plus an encounter with state, context, reproducibility, permissions and observability.

Other sessions attack different failure surfaces. Estuary’s change-data-capture lab treats freshness as correctness: participants capture live database updates and materialize them transactionally before asking how current the warehouse really is. Airbyte’s workshop goes from a custom source through MCP to a natural-language agent, then requires participants to structure context and measure answer accuracy. A Braintrust lab moves downstream to production traces, where teams must find recurring failure patterns rather than manually inspect every run.

Those are stronger acceptance criteria than “the agent answered my demo question.” Together they ask whether inputs arrive atomically, whether context is structured, whether permissions survive automation, whether runs can be reproduced and whether failures are observable.

Action raises the control bar

The program also pushes beyond read-only chat. Sigma’s lab has participants build an agent that can take action, while Prefect’s FastMCP exercise exposes tools and resources through a hosted server that reserves a shared gift. Vercel’s session connects an analytics agent to enterprise context and communication surfaces including Slack and iMessage.

That progression changes the risk model. A wrong chart is bad; a wrong action can alter shared state. The workshop descriptions do not constitute independent proof that these systems are safe, scalable or accurate, and MotherDuck is promoting its own event and platform. But the curriculum is still a useful signal of what practitioners are being asked to build and debug.

The practical checklist is clear: do not evaluate a data agent only on generated SQL. Test the freshness and transactional integrity of its inputs, force upstream breakage, measure answers against known cases, preserve permission boundaries, retain traces and rehearse recovery before enabling actions. The model is one component. The production system is the product.