Run coding agents like a medical team to fix bugs
Between a Wednesday evening and a Friday afternoon in April, MOLT, the tool Cockroach Labs ships for migrating databases to CockroachDB, gained the ability to migrate from IBM Db2 sources. Task: Cockroach Labs used a custom agent pipeline called MOLT Sinai to add IBM Db2 support to their MOLT migration tool. A planning agent decomposed a single GitHub issue into 32 sub-issues with dependency graphs, while coding agents generated schema converters, fetch/verify paths, and test fixtures. Human 'Chiefs of Medicine' reviewed escalated issues and approved merges, ensuring production-grade quality through mandatory plan reviews and I-PASS handoffs.; tools: custom-built or unnamed agent. Human role: The original post does not say how a person stayed involved.; limitations: The workflow requires a complex GitHub Actions infrastructure with role-specific prompts and strict labeling conventions. It is optimized for high-stakes database code where correctness outweighs speed, making it unsuitable for rapid prototyping. Human approval is mandatory for merges in production mode, and the system suffers from 'bureaucratic' delays where agents may escalate trivial issues or require multiple rework rounds for minor deviations..
Tools
custom-built or unnamed agent
Published on Agent Work Atlas:
Overview
Between a Wednesday evening and a Friday afternoon in April, MOLT, the tool Cockroach Labs ships for migrating databases to CockroachDB, gained the ability to migrate from IBM Db2 sources.
Workflow
Cockroach Labs used a custom agent pipeline called MOLT Sinai to add IBM Db2 support to their MOLT migration tool. A planning agent decomposed a single GitHub issue into 32 sub-issues with dependency graphs, while coding agents generated schema converters, fetch/verify paths, and test fixtures. Human 'Chiefs of Medicine' reviewed escalated issues and approved merges, ensuring production-grade quality through mandatory plan reviews and I-PASS handoffs.
Reported result
The Db2 support feature was completed in under two days with zero human-written code, costing $4,172 in tokens compared to $160K for the previous Oracle implementation. The pipeline merged 27 pull requests and achieved test coverage equivalent to PostgreSQL. However, the system generated significant bureaucratic overhead, including 55 rework cycles and 9 escalations, demonstrating that quality assurance comes at the cost of speed and increased process complexity.
Human role
The original post does not say how a person stayed involved.
Limitations
The workflow requires a complex GitHub Actions infrastructure with role-specific prompts and strict labeling conventions. It is optimized for high-stakes database code where correctness outweighs speed, making it unsuitable for rapid prototyping. Human approval is mandatory for merges in production mode, and the system suffers from 'bureaucratic' delays where agents may escalate trivial issues or require multiple rework rounds for minor deviations.
Try it
Act as a Chief of Medicine for a coding agent pipeline. I have a GitHub issue [replaceable: issue URL or description] requesting [replaceable: feature, e.g., 'add support for X database']. 1. Analyze the issue and decompose it into sub-issues with a dependency graph if it exceeds 1,000 lines of estimated code. 2. Assign a 'Fellow' agent to create a treatment plan (diagnosis, files to change, tests) for the first sub-issue. 3. Review the plan as a 'Review Attending'; approve or reject based on scope and risk. 4. If approved, instruct the agent to implement code and tests. 5. Verify that tests fail before the fix and pass after. 6. Check for I-PASS handoffs if the agent gets stuck. 7. Do not merge until a human confirms the PR diff and CI status. Start with a dry run on a small, isolated module to validate the workflow before proceeding to the full feature.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.