Loading today’s stories
Build1 distinct publisher3 min readPublished
MotherDuck's first acquisition buys the runtime that was already executing its agent-written pipelines. The instructive part is what Jordan Tigani says the agents cannot do.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The scarce input in agent-built data engineering is not the code. Jordan Tigani's account of why MotherDuck changed course says so plainly: the company had first gone looking for a third-party tool it could recommend to customers as an easier way to get data into its warehouses [13], and then, in his words, "When AI suddenly started to be able to solve data problems, we realized we were thinking about the problem wrong" [14]. The generated connector stopped being the hard part. Somewhere to execute it safely, holding credentials, running on a schedule, became the hard part [16].
Tower had been selling that exact description of the work before the agents showed up. Serhii Sokolenko, who founded the company in Germany in late 2024 with fellow ex-Snowflake engineer Brad Heller, told The New Stack in March that packaging code, deploying it to the right infrastructure, wiring up credentials and maintaining it is data engineering's "last mile" [4][5][6]. Tower's own browser agent, Tower Control, takes a plain-language request through to a deployed app that runs [7], which it could only do because it owned the runtime underneath.
That is also why MotherDuck could ship Flights "in only a matter of weeks", by Tigani's account [17]. Borrowed foundations buy speed and then charge for it: a feature that ships fast on another company's execution layer is a feature you cannot re-price or re-architect on your own timetable. Tigani is candid about the alternative, saying it was tempting to build in-house but that trying Tower surfaced "a bunch of problems we were going to have to solve to make our underlying infrastructure actually work well" [19].
Some arithmetic the coverage leaves implicit. MotherDuck was founded in 2022 and this is described as its first acquisition in a four-year history [1][8], which places the deal around 2026, roughly two years after Tower was founded [21]. A company that has raised about $100 million and built its identity on DuckDB queries running on a laptop [8][9] has concluded that the piece it cannot rent is the one that runs other people's generated Python. The line The New Stack put in its headline, that you can rent a feature but you cannot rent a foundation [20], is carrying more weight than slogans usually do: to the customer, Flights and Tower were one product; to MotherDuck's balance sheet and roadmap, they were two companies.
The competitive shape follows from the plumbing. MotherDuck's agent surface is MCP, the same channel it uses to let agents interact with data directly [10], and the Flights runtime was exposed through that same MCP server [11]. Any warehouse offering agents that write and schedule jobs needs a sandbox and a scheduler behind that interface, whoever's name is on it. For a buyer, the question worth asking has moved off which model writes the pipeline and onto whose infrastructure holds the warehouse credential while an unreviewed job runs unattended at three in the morning. MotherDuck has just changed its own answer from "a partner" to "us".
Ranked by verification strength, evidence, and original report placement.
MotherDuck announced on Tuesday its acquisition of Tower, the first acquisition in MotherDuck's four-year history, bringing Tower's technology and team in-house as MotherDuck pushes further into AI agents that build and operate data pipelines.
Tower's technology was already powering MotherDuck's AI-built data pipelines before the acquisition.
Tower is a managed runtime for Python pipelines: it packages the code, deploys it, and keeps it running in production.
Tower was founded out of Germany in late 2024.
Tower's founders are ex-Snowflake engineers Serhii Sokolenko (CEO) and Brad Heller (CTO).
Sokolenko told The New Stack in March that once a developer or an AI assistant has written pipeline code, someone still has to package it, deploy it to the right infrastructure, wire up credentials and maintain it, which he called data engineering's "last mile".
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single-publisher vendor account, verifiable on the deal but thin on substance
The core facts — that the acquisition happened, that Tower was already the runtime under Flights, what Tower is, who founded it — are stated directly and consistently by a named trade publication drawing on the acquirer's CEO and its own earlier interview with Tower's CEO. But there is exactly one source in the cluster, one named speaker on the deal, no deal terms, no metrics, and no independent or adversarial verification of the operational claims about sandboxing, scheduling or the weeks-long integration.
Real production dependency, no usage magnitude
Adoption is demonstrated in kind but not in size: Tower's runtime was already executing jobs created and scheduled inside a shipped MotherDuck feature (Flights, launched in June), and MotherDuck describes itself as Tower's largest customer. That is genuine production use by one significant customer. Nothing in the sources quantifies end-user adoption of Flights or Tower Control, names another Tower customer, or reports job volumes, so broader market uptake cannot be scored higher.
Mildly overstated: agentic framing outruns disclosed substance
The story is slightly forward-leaning relative to what it shows. 'Prompt to production', a headline aphorism supplied by the buyer, and unquantified speed claims sit above a disclosure base with no price, no metrics and no third-party verification. The gap stays small because the same article deflates the strongest agent claims itself — Tigani's point is precisely that Claude writes connectors but cannot sandbox or schedule, which is a limits-first framing rather than a capability boast.
Acquirer-driven announcement with access-based sourcing
Every substantive quote comes from the buyer's CEO on the day of his own announcement, and the narrative he supplies — own the foundation, we shipped in weeks, we became their largest customer — serves MotherDuck's positioning against cloud-first incumbents. The publisher has its own access incentive, having interviewed Tower's CEO in March and now carrying the follow-on exclusive-style piece as the headline quote. No adversarial or customer-side voice is present to offset this.
Confident on the event, weak on consequences
Confidence is high that the acquisition occurred, that Tower is a managed Python pipeline runtime, and that it underpinned Flights — these are specific, named, internally consistent statements from a known trade outlet. Confidence is low on anything about scale, price, durability, security posture, or what happens to Tower's remaining customers, because the cluster has one source, one interested speaker and no metrics.
build
Anthropic's Browser Use hands Claude element refs, and hands you the browser1 distinct publisher
product
Thomson Reuters spent $40M to make a $450K training run worth doing2 distinct publishers
build
DuckDB is growing a server, and someone on your team will have to run it1 distinct publisher
build
Databricks says the hard part of warehouse migration was the stored procedures, not the data1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.