Cockroach Labs通过构建名为MOLT Sinai的AI编码代理系统，采用教学医院模式管理高风险代码生成工作，在不到两天内完成了IBM Db2数据库迁移支持，仅耗资4172美元1。这一成绩相比2024年传统方式处理Oracle支持的9个月开发周期和16万美元工程成本，实现了164倍的加速和38倍的成本节省1。
该系统在2024年4月21日至9月11日的5个月期间内消耗了13.5万美元的Claude token，平均每个issue的成本为84美元1。在此期间，系统共提交了1238个拉取请求，其中约1299个issue由系统自动生成，占比接近50%，共生成超过100万行代码1。Db2支持工作具体产生了32个子issue、27个合并的拉取请求、55次代码评审反馈和9个议题升级1。
该系统遵循教学医院模式的核心规则运作：无计划前不编码、计划必须经过评审、禁止跳过测试、遇到问题时提交标准化交接单据（I-PASS）1。系统维护的技能文件共包含100,000词，其中23%（约23,000词）在审计后被判定为可移除的冗余内容1。
Cockroach Labs has deployed an AI coding agent system called MOLT Sinai that applies teaching hospital principles to manage high-risk code generation work. Operating under this framework from April 21 to September 11, 2024, the system completed IBM Db2 database migration support in less than two days at a cost of 4,172 dollars, consuming 135,000 dollars in Claude tokens across the five-month period while generating over one million lines of code across 1,238 pull requests. 1
The accelerated timeline and cost efficiency represent a dramatic improvement over traditional approaches. The IBM Db2 support effort, which required no manual coding, contrasts sharply with Oracle support work completed in 2024 that took nine months and incurred 160,000 dollars in engineering costs—reflecting a 164-fold speedup and 38-fold cost reduction. 1 The Db2 work produced 32 sub-issues, 27 merged pull requests, 55 rounds of code review feedback, and nine escalated issues. 1
The teaching hospital model enforces strict operational rules: no coding without a plan, all plans must be reviewed, tests cannot be skipped, and when blocked, engineers must submit standardized handoff documentation using the I-PASS format. 1 The system maintained skill files totaling 100,000 words, though audits identified that 23,000 words—representing 23 percent of the total—contained redundant content deemed removable after review. 1 Approximately 1,299 issues were automatically generated by the system, accounting for roughly 50 percent of the work, with an average cost of 84 dollars per issue across the five-month span. 1
评论
还没有评论，欢迎留下第一条。