软件工程师Steve Yegge分享了基于Claude Fable 5构建的Wheelhouse多智能体编排系统的实践经验。[1]该系统用于加速其开发25年历史的MMO游戏Wyvern的开发。[1]Wheelhouse包含18个Fable命名智能体和更多Opus 5工作智能体，配合Beads知识图谱和Claude账户循环实现多角色编排。[1]
系统运营规模庞大。[1]月度API消耗约87,000美元（69亿tokens，96%缓存命中率），通过12个额外Max账户和账户轮转的成本控制方案，将实际月度支出降至2,800美元。[1]每日平均产生175个"实际"提交，峰值达250个。[1]Beads系统日均产生12,000个Git提交，支持依赖、父子、触发等关键编排原语。[1]
Yegge发现传统持续集成/部署方案在高频提交下失效，转而采用"Land Rush"方案：当提交队列达到100条时放弃传统二分查找，直接批量合并120-150个提交后进行群体诊断修复。[1]项目知识库包含约100个Markdown文件和30个私有Skills，用于编码组织知识并降低token消耗。[1]系统还引入多个角色智能体进行生产运维，包括Gargoyle（SRE）、Warden（滥用监控）、Sheriff（fleet管理）等。[1]
Yegge预测，"人工代码审查在明年完全消失"，并认为传统CI/CD模式因智能体倍增的提交率而必然崩溃。[1]
Software engineer Steve Yegge has unveiled a novel approach to accelerating development on his 25-year-old MMO game Wyvern by building Wheelhouse, a multi-agent orchestration system powered by Claude Fable 5.[1] The system combines a Beads knowledge graph, Claude account rotation, and multiple specialized agents to handle the exponential increase in code submissions generated by AI-assisted development.[1]
Wheelhouse operates with 18 Fable-named agents and additional Opus 5 work agents, processing an average of 175 "real" commits daily, with peaks reaching 250.[1] Traditional continuous integration approaches proved inadequate under this submission volume, leading Yegge to implement a "Land Rush" strategy: when the merge queue reaches 100 pull requests, the system abandons sequential validation and instead batches 120 to 150 commits for simultaneous integration followed by group diagnostics and repair.[1] This approach mirrors practices already adopted in game development operations, though arrived at independently.[1] The Beads system itself generates approximately 12,000 Git commits daily, supporting dependency resolution, parent-child relationships, and trigger-based orchestration primitives.[1]
Cost management has been achieved through aggressive optimization: despite initial monthly API consumption of approximately $87,000 (representing 6.9 billion tokens with a 96% cache hit rate), account rotation across 12 additional Max-tier accounts has reduced operational costs to $2,800 per month.[1] The project maintains institutional knowledge through roughly 100 Markdown documentation files and 30 private Skills, designed to encode organizational practices and reduce token consumption.[1] Beyond development acceleration, Yegge has introduced specialized production agents including Gargoyle for site reliability engineering, Warden for abuse monitoring, and Sheriff for fleet management.[1] Yegge predicts that "human code review will completely disappear next year," arguing that traditional CI/CD architectures will necessarily collapse as AI agents generate submissions at unprecedented scale.[1]