Build It Yourself: What You'll Make at the Data Outpost Workshops

retained text

Build It Yourself: What You'll Make at the Data Outpost Workshops
· 10 min read
Reading about agents is fun. Building one with your own hands is better! On Wednesday, November 4, Data Outpost kicks off with a full day of hands-on workshops in San Francisco. You'll build agents, wire up the data they need, and learn what it takes to run them in production. Bring a laptop, because most of these sessions end with something running on it.
There are 11 workshops across 3 time slots (9:15 AM, 11:00 AM, and 2:30 PM), and they run in parallel, so you'll pick 1 per slot. They're at Convene 100 Stockton, the same venue as the main day of talks and panels on Thursday, November 5. (Here's what you'll learn on day 2.)
Sound good? Register for Data Outpost here! Need help choosing? Read on!
Here's what you'll build in each workshop, plus my take on each one.
Build Your Own Agent
4 workshops where you walk out with an agent of your own: one you can message on Slack, one that builds a whole data stack, one that lives in your BI tool, and an MCP server that reserves you a gift.
Build a Data Agent with eve
Colton will walk through the fundamentals of the eve framework and some popular use cases, then help you build your own personal data analytics agent. You'll connect it to the context sources in your enterprise (such as MotherDuck) and talk to it wherever you already work: Slack, iMessage, or embedded in your website.
My take: I love that this is open source! They have thought through the tough problems in the space: durable execution, sandboxing, and more.
The Next Data Stack Has No UI: Building End-to-End Analytics with Agents and Ducks
What if your next data platform is a conversation instead of a pile of dashboards and config screens? You bring an agent (Claude Code or similar), Mehdi will bring some raw data and coach you through the process. You'll point your agent at those files with DuckDB, then let it profile the data, design a schema, ingest it, write the transformations (complete with data quality checks), and publish to MotherDuck.
Along the way you'll break the data on purpose to see how to build a resilient system, and talk through where the DuckDB ecosystem shines for agents and where it falls apart: state, context, reproducibility, permissions, and observability. You'll leave with a working end-to-end stack, a repo of prompts and guardrails, and the experience of building a modern platform from scratch.
My take: If you use AI to speed up just 1 small part of a process, you don't see much benefit. This is all about applying AI to the entire process end to end, and it will speed things up dramatically. In my experience, finding and ingesting data takes the longest, and this will help!
See it live: a hands-on Sigma + MotherDuck lab
Workshop: Zalak Trivedi (Product Leader, Sigma) and Tyler Johnson (RVP, Solutions Manager, Sigma) · 2:30 to 4:30 PM. On the agenda
Most BI demos show you a finished dashboard. In this lab you build one, with Sigma connected directly to a live MotherDuck warehouse. The first half is classic analysis: drag fields into tables and build a pivot table from scratch.
In the second half you'll put Sigma's AI to work. Ask Sigma Assistant questions in plain language, then build your own Sigma agent and watch it take action on your behalf.
You'll leave with a working workbook connected to MotherDuck and a clear sense of what your team could build in an afternoon.
My take: My favorite part of this is the "take action" piece. That's where the benefits of agents really compound.
Build an MCP Server and Unwrap a Mystery Gift with FastMCP and Horizon
Workshop: Edward Park (Engineering Director, Prefect) and Radhika Gulati (Product Marketer, Prefect) · 2:30 to 4:30 PM. On the agenda
Yes, there is an actual mystery gift! You'll use FastMCP and Horizon (Prefect's MCP hosting platform) to build an MCP server, expose tools and resources, and take it from local development to a hosted deployment. Then you'll connect it to a shared gift service, so an agent can check which numbered gifts are still available and reserve one for you.
You'll leave with a working hosted MCP server, hands-on experience connecting agents to a shared service, and a gift to take home.
My take: Building an MCP is a great way to scale your expertise across the company.
Get Data Agent-Ready
An agent is only as good as the data it can reach. These 5 workshops get your data in shape: migrated from your current warehouse, stored in an open lakehouse, streamed fresh from a live database, synced from a custom source, and pulled out of messy documents.
What to move, what to fix, what to leave behind: migrating to MotherDuck
Thinking about moving to MotherDuck, but dreading peeling back the duct tape on your current setup? Dumky will help you build your own system for migrating data: an inventory built from query history, logs, and access patterns, the right strategy for each workload, and the data movement, transformations, and consumption layers to match.
It applies whether you're coming from Snowflake, BigQuery, Postgres, or something else, and whether your data feeds dbt, Hex, Power BI, or a customer-facing web app. You'll leave with the runbook to migrate with confidence.
My take: The hardest part of a migration is always in the details. Making sure you handle all workloads and can validate that the data is correct will really speed things up!
Lakehouses for Robots: Scaling DuckLake and Iceberg for Agent Workloads
Workshop: Alex Monahan (MotherDuck) and Matt Martin (Staff Engineer, State Farm) · 9:15 to 11:00 AM. On the agenda
Data lakehouses are open, scalable, and interoperable, but most are not ready for agents. We’ll show you how to query Iceberg and DuckLake using both DuckDB and MotherDuck with the agent of your choice. You’ll also learn how to take full advantage of your local laptop so you can build even faster.
Whether you are new to lakehouses or a seasoned expert, you’ll end the session with a working lakehouse example ready for the flood of agents.
My take: I'm a bit biased - I'll be leading this one with Matt! I really like the lakehouse architecture, so we'll build one that's agent ready.
From Scheduled Syncs to Real Time Data Using Change Data Capture: A Hands-On Lab
Stale data used to be tolerable, because the person reading the dashboard knew it was a few hours behind. Agents query constantly and act on what they find, and a partially arrived transaction looks exactly like an arrived one. That makes stale data a correctness problem - your agent will draw the wrong conclusion!
In this guided lab you'll capture updates from a live source database as they happen, materialize them into your warehouse in a transactionally safe way, and answer how current your data actually is.
My take: Nobody wants to point agents at their production database. Replicating your data to a safer and faster environment for agents is a great plan, and Estuary can help make that seamless!
Turning Your Data Into AI Context, Powered by Airbyte
Go from a raw source to an agent that answers questions about your data. You'll build a custom source in the Airbyte Cloud connector builder, use the Airbyte MCP to sync it into MotherDuck, then build an agent on top that explores the data in natural language and puts a dashboard on it.
Then comes the important part: structuring the agent's context so its answers are consistently correct. You'll leave with a working connection, a dashboard on your synced data, and an agent whose accuracy you've actually measured.
My take: I've pulled data out of all kinds of legacy data systems (in a past role), so it is really valuable to be able to customize your data ingestion. I'm certain Airbyte + agents make this way more fun than it used to be!
From Unstructured Documents to Agent-Ready Data
Workshop: Palak Agarwal (DevRel Lead, Reducto) and Josh Nkoy (Founding Solutions Engineer, Reducto) · 2:30 to 4:30 PM. On the agenda
PDFs, scans, images, and all kinds of messy documents go in, and a structured, searchable archive comes out. You'll use Reducto's MCP to build a curator, then pick your own collection and create a mini-exhibition: a poetry show made from drafts and annotations, a moment in history told through letters and newspapers, or a championship season rebuilt from box scores and scouting reports.
You'll leave with a working document pipeline and an exhibition you can share.
My take: A scanned in PDF is the final boss of messy data, but sometimes that's where the core nuggets of organizational context live.
Run Agents in Production
Shipping an agent is the start. These 2 workshops cover what comes after: finding out what your agent actually does in production, and what it takes to serve lots of long-running agents at once.
The Agent Loop: From Traces to Patterns to Better Agents
Workshop: Izzy Hurley (Eval Engineer, Braintrust) and Spencer Seale (Solutions Engineer, Braintrust) · 11:00 AM to 1:00 PM. On the agenda
Your agent is in production. Now you need to understand what keeps going wrong. Agents rarely fail in obvious ways, and no team can manually review every production trace.
You'll run the full evaluation loop twice: first with a physical building challenge that makes the ambiguity tangible, then with a real AI agent. You'll explore production traces, uncover recurring patterns, build an evaluation dataset and scorer, and test an improvement.
You'll leave with a repeatable workflow for turning real agent behavior into more reliable AI.
My take: If you want your company to trust AI, you need evals. The folks at Braintrust have a ton of expertise here!
LLM Inference for the World of Agents
Learn how your agent works, down to the details, so you can build or customize your own.
An agent might run for minutes, make 100 tool calls, and generate tens of thousands of tokens before it's done. That's a very different workload from the single prompt and response that most examples describe.
Zain will take apart a coding agent (think Claude Code, Codex, or Cursor) into its model, harness, context, tools, and memory, then follow the tokens into the inference engine: tokenization, continuous batching, prefill and decode, KV-cache management, prefix caching, and more.
You'll leave with one continuous mental model of the entire agent stack.
My take: We used to design software, but I think that we will increasingly be designing the systems we use to build software. This is a great way to get comfortable with the agent-centric future we're headed to.
See You at the Outpost!
Every workshop on this list ends with something you can keep using: an agent, a pipeline, a runbook, or a mental model. And they share the big idea from the main day: agents get better when the data underneath them is done well.
A quick heads up: workshops run in parallel, so pick 1 per time slot. Then stick around for Thursday, November 5, for a full day of talks and panels on the data stack after agents, AI at scale, semantic layers, and more.
Register for Data Outpost and bring your laptop.
See you in San Francisco!
Table of contents
Read Next

document fetched 2026-10-24 10:06