Five months treating bugs like patients and coding agents like a medical team

A software pipeline was built that treats automated coding work like a teaching hospital: issues become “patients,” merges are “discharges,” and specialized agent roles (Triage Nurses, Fellows, Review Attendings, Discharge Nurses, Charge Nurse, Infection Control, Safety, Research) run as GitHub Actions that apply and remove labels to drive workflow stages. The system enforces strict safety rules - no code before a reviewed plan, no improvisation, tests cannot be weakened, structured I-PASS handoffs when stuck, and a discharge audit that verifies reviews and CI - plus an append-only precedent log so agents consult past human decisions before escalating. The design prioritizes correctness over throughput because migration tools must preserve customer data integrity.

Over five months the hospital pipeline generated and merged over a million lines of code while keeping high quality: Db2 migration support was implemented in under two days with no human-written code and a token cost of $4,172, compared with a prior Oracle effort that took nine months and cost about $160K (roughly 164× faster and 38× cheaper). The work produced 32 sub-issues, 27 merged PRs, 55 review-driven send-backs, nine escalations (two to humans), and seven total reverts; test coverage matched the best-tested dialect. The experiment shows structured, role-based agent workflows can safely drive complex engineering work at scale, trading bureaucracy and latency for enterprise-grade reliability.