GitHubWhat they did
Adam Storm and Rafi Shamim at Cockroach Labs built a pipeline they call MOLT Sinai, where each agent holds a role borrowed from a teaching hospital: a planner that reads a GitHub issue and splits it into sub-issues with a dependency graph, implementers, reviewers kept in separate contexts, a discharge nurse that checks the review actually happened before merge, a charge nurse that does rounds every thirty minutes, and a precedent log of decisions made by the human Chief. Issues are patients and merging is discharge. The first large job was a single issue asking for IBM Db2 as a migration source: a schema converter, fetch and verify paths, a vendored ANTLR grammar, a CI Docker image and test fixtures. The planner cut it into fifteen sub-issues, two of which were split again, and the agents filed a dozen more issues against their own work. The pipeline ran from April to September 2026 on a mirror of the MOLT repo; after its first week they switched on a mode where a human must approve every merge. It started on Opus for every stage and now plans on Fable, which picks the model for implementation.
How it went
Db2 support took under two days and $4,172 in tokens: 32 sub-issues, 27 merged pull requests, 55 review send-backs and nine escalations, two of them to a person. They compare that with roughly nine months and $160K of engineering time for Oracle support in 2024. Over five months the pipeline consumed just over $135,000 in Claude tokens, about $84 per issue, with issues taking one to two days, mostly waiting on human review. The framework repo has about 1,570 commits, 85 percent of them written by the pipeline, and seven merged changes had to be reverted. A July audit found 23 percent of the 100,000 words of skill files redundant. Their Migration Assistant product, now in preview, was almost entirely written this way. They have no long-term maintainability data yet.
Worth knowing
Their first mistake was auto-merge: output arrived faster than anyone could verify, and sibling sub-issues quietly dropped work each assumed the other would do, which they fixed with a deferred-scope ledger. The pipeline is internal and not public, though open-sourcing has been discussed. HN commenters compared it to OpenAI's Symphony and several described tearing out similar agent factories that generated more process than value. The authors say it is a side project for both of them.
