Cockroach Labs shipped MOLT's IBM Db2 migration support in two days via agent-run pipeline

Cockroach Labs' database migration toolset MOLT can now move data off IBM Db2, and according to a company engineering blog post that reached the front page of Hacker News, essentially none of that code was written by a person. The feature went from a single GitHub request to merged work between a Wednesday evening and a Friday afternoon in April, with a total token bill of $4,172.

What shipped

Db2 is one of the earliest commercially available relational databases, with a rich SQL dialect, a complicated type system and its own wire protocol. Supporting it in MOLT required a new schema converter, a new Fetch path for pulling rows into CockroachDB, a new Verify path for comparing data once loaded, a vendored ANTLR grammar, a Docker image for CI, and more than ten thousand lines of test fixtures.

For comparison, Cockroach Labs writes that adding Oracle support to the same tools in 2024 took about nine months and roughly $160,000 of engineering time. By the company's own accounting, the Db2 effort came out 164 times faster and 38 times cheaper, with no human writing the code.

Inside MOLT Sinai

The work ran through an internal system Cockroach Labs calls MOLT Sinai, a pipeline modeled on a teaching hospital and run on GitHub Actions. Issues are treated as patients, merging as discharge, and the humans in charge act as Chiefs of Medicine. State lives in issue labels and scratch files: applying a label triggers a workflow that loads a role-specific prompt, and each agent appends a structured note to a per-issue chart that later agents parse.

The cast mirrors a hospital ward. A planning agent decomposes large requests into smaller ones with an explicit dependency order. A Fellow agent must reproduce a problem, narrow down a diagnosis, and post a treatment plan covering files to change, tests to add, risks and open questions before touching code. A separate Review Attending, prompted to find fault rather than fix, approves or rejects that plan first. A Charge Nurse checks every thirty minutes for stalled work, an Infection Control agent locks down every workflow if the main branch breaks, and Safety and Research departments run process reviews and propose new work on their own schedules.

A few rules do most of the safety work, according to the post: code is only written after a reviewed plan, scope changes force a new plan rather than improvisation, and agents may not disable, skip, weaken or modify a test to make it pass. A failing test is presumed correct unless proven otherwise.

Quality gates over throughput

Cockroach Labs frames the whole design as a bet on quality rather than speed. The company notes that other recent agent-coding efforts, such as parallel agent swarms with merge-handling layers or a fast experimental SQLite rebuild, optimize for how quickly code gets written. That is the wrong objective for migration tooling, the post argues, because a subtle bug could corrupt customer data on its way into a database whose core promise is correctness. The team says it would rather let agents run ten times longer than land bad data in a customer migration.

The Db2 ticket illustrates the process. The planning agent judged the request too large for one pass and split it into fifteen sub-issues ordered by dependency: foundation first, then the type system, the row iterator, Fetch, Verify, Convert, CI and test data. Two sub-issues were split again, and agents filed roughly a dozen additional issues against their own output, covering fixture gaps, a type-mapping bug and an isolation-level fix. By the time the parent issue closed, thirty-two sub-issues had been opened, twenty-seven pull requests had merged, review had sent work back fifty-five times, and nine issues had been escalated, two of them to a human. Cockroach Labs reports that test coverage matched its PostgreSQL dialect, its best-tested.

The hospital has now been running on real work for five months, and the company acknowledges the metaphor brings familiar downsides: bureaucracy, wait times and cost.

Why it matters

For CockroachDB, Db2 support opens a path from one of the oldest and most entrenched enterprise database ecosystems, a demographic that rarely moves quickly and evaluates tooling on trust.

The broader signal is the governance model. The headline numbers, two days and $4,172 against nine months and $160,000, are striking, but the durable lesson in the post is what sat between the agents and the main branch: mandatory plan review, escalation paths to humans, and hard rules against tampering with tests. As coding agents proliferate, the differentiator for production-quality results may be process design like this rather than raw generation speed.