Want to know how our product team uses MotherDuck to build with agents? Join Sheila Sitaram (Product Lead for Compute and Agent Workloads) and Zoey Weaver to talk through things like: ​🦆Sorting customer call notes and support tickets into themes with prompt_jev(), so the same feature requests surface every week ​🦆A scheduled agent that pulls context from Slack, support tickets, and Linear to write the weekly business review analysis ​🦆Dives that show how customers' agent workloads use compute Grab a spot here:

From 17.5% to 99.7% accuracy on hard data analysis questions. That’s the difference Guides made in our DABStep benchmark evaluation. Gemini 3 Flash's gain with Guides across 419 held-out questions: > Hard: 17.5% → 99.7% (+82.2 percentage points) > Easy: 78.9% → 100% > Overall: 27.9% → 99.8% The difference comes down to Guides - the agent could retrieve the business context it needed while working through each question. It works because your schema doesn’t tell the whole story. Metric definitions, join logic, deduplication rules, and business caveats all shape the right answer. Your agents 𝘯𝘦𝘦𝘥 that information if they're going to provide correct answers. Guides puts that knowledge in MotherDuck, alongside your data, for your agents. Write it once as markdown, and agents connected through our MCP server can find and load the context each question needs. Give your agents a Guide:

Ducktober, Drop 03: Giga 🦆 One very big duck. Tens of terabytes. No cluster. Giga is MotherDuck’s largest instance: roughly the CPU and RAM of a Snowflake 3XL, in a single machine. On the 10 TB TPC-DS benchmark it ran all 99 queries at a 5.6-second geomean. No nodes to size. No cluster to tune. When the data gets big, scale up. Catch up on Ducktober and read our CEO's recent blog on Giga→ Five more to go. The next one comes through tomorrow at 9am PT.

If you're a duck, you redact data in the only way you know how: with a paintbrush. Scary! For humans and agents, MotherDuck now supports Row & Column Level Security as well as Data Masking. Learn more:

Sometimes a wizard needs a castle of their own. Bring Your Own Compute: a private MotherDuck region in your own AWS account. Happy Ducktober. Learn more here:

DuckDB Monthly #46# is out. Spotlight: @marklit82, of 1.1 Billion Taxi Rides fame, who has been showing what DuckDB can do for GIS since 2023. Plus Jev typed AI answers in SQL, DuckDB 2.0's PEG parser and async S3 reads, and DuckLake maintenance.

Today is DuckDB 2.0 day! 🦆🎉 Congrats to the DuckDB team and everyone who has contributed. We're super excited for async I/O, a first-class VARIANT type, triggers, and a lot more. For MotherDuck users: our release and support come a few weeks after major versions of DuckDB. This is because we're working to make sure everything is stable and scalable for all customer workloads as DuckDB rapidly evolves its engine with every release. Read the full announcement from the DuckDB team: Join us on Oct 29 for a livestream with Alex Monahan diving deep into our favorite parts of “Cinnamon Teal”. Save your spot ->

DuckDB 2.0? Join us on 10/29 to walk through what's new, how it's impacting workload speeds, and lots more with Mehdi Ouazza. In his benchmarks, reading a 2.2 GB Parquet file from S3 dropped from 18.8 seconds to 7.7 seconds, and a deep recursive CTE went from several seconds to a tenth of a second...

Mike Failor calls Byzfunder's regex rules the secret sauce. Thousands of lines of it, built up over years to label every transaction on a small-business applicant’s bank statements. But there was no practical way to check them. The label that matters most is a payment from another lender. A business already carrying competitor loans is a riskier one to fund. When the regex rules have never seen that lender, the loan slips through as an ordinary transaction. On his first day with Jev in MotherDuck, Mike (Byzfunder’s Chief Risk Officer) ran 120,000 transactions through an audit the team had never done before. It took under a minute. The check is a SQL query. Jev reads each transaction and scores how likely that transaction is a payment from another lender. An analyst checks the transactions where that score disagrees most with the regex. Every confirmed miss gets fixed in the regex, so the next check has less to find. “With AI running right inside MotherDuck, we're catching competitor loans in bank statements we would have otherwise missed. Closing those gaps gives us a more accurate assessment of every business we lend to.” Full case study:

A week from today, join Alex Monahan and Matt Martin to cover all things DuckLake performance optimization. Their new chapter is live, so they'll cover how to: ​🦆Find the bottleneck first with metadata logging, EXPLAIN ANALYZE, and JSON query profiles (and why handing those profiles to an LLM works well) ​🦆Pick partition keys and sort orders so queries skip whole files and row groups, and spot when bucket partitioning works against you ​🦆Sort on two columns at once with space-filling curves, something the Iceberg spec doesn't allow ​🦆Tune compression, file size, and row group size together instead of one at a time ​🦆Keep the catalog and storage lean with inlining, compaction, snapshot expiration, and cleanup, and know which of these actually make queries faster Register here to join them:

Incoming dispatch from the edge 📡 Data Outpost is 2 days for builders at the forefront of AI + data, across 4 themes: Systems - Agentic data infrastructure Product - Interfaces and workflows Trust - Memory, provenance, identity Craft - Build with taste Nov 4-5 in San Francisco, with hands-on workshops on day 1 and talks on day 2. 🎟️ Tickets:

Weekend plans? Connect your favorite agent app to MotherDuck MCP and see what you can build with Flights, Dives and the MCP connector. Available on Claude, Cursor, ChatGPT, Copilot and more.