ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-10-27 17:04

WorkOS软件工厂

3 steps1 pieces5 Sources

美国企业基础设施公司 WorkOS 的产品工程师 Ryan Cooke 在2026年 AI Engineer World's Fair 上提出,当前行业对“软件工厂”的定义存在误区:仅依靠沙箱和模型生成代码并非真正的自动化,而是将工程流程本身编码进系统。

【背景】Cooke 指出,Ramp、Spotify 和 Stripe 等公司虽已建立类似机制,但普遍以 PR 数量和 AI 代码占比为衡量标准,这混淆了“生产速率”与“价值交付速率”。WorkOS 的初始实验显示,简单的沙箱方案产出的结果与工程师本地运行编码工具无异。

为解决这一问题,WorkOS 构建了名为 Horizon 的基础设施编排层和 TARS 交互界面。其核心在于通过内部 MCP 网关连接 Slack、Linear 和 GitHub,利用工单依赖关系和完成 Webhook 自动触发后续任务,并将规划过程(如 Hilltop 文档拆解)自动化。这一转变旨在让 AI Agent 参与工程协调而非仅生成代码,从而提升功能交付速度并降低缺陷率。

How it got here

  1. 2026-05-06

    WorkOS 发布博客《The self-driving codebase》,详细介绍内部项目 Horizon,阐述如何通过事件驱动代理解决编排瓶颈,并提及受 Ramp Inspect 和 Stripe Minions 启发。

    workos.com ↗
  2. 2026-08-17

    Mastra CTO Abhi Aiyer 在 Agent Night 演示开源软件工厂,强调规则引擎和人工审查环节的重要性,指出自主代理目前仍缺乏对公司知识库的理解。

    workos.com ↗
  3. 2026-10-27

    Ryan Cooke 在 AI Engineer World's Fair 发表演讲“No, That's Not a Software Factory”,论证行业正在优化错误的变量,展示 WorkOS 如何将规划、协调和上下文编码进自动化流程。

    youtube.com ↗

What is still unsettled

  • WorkOS 的 Horizon 系统在引入自动化后,实际的功能交付时间(Feature delivery time)和缺陷率相比传统流程有何量化改善?
  • 其他采用类似模式的公司(如 Ramp、Stripe)是否也转向了以结果为导向的指标体系,还是仍停留在输出指标?
  • Mastra 提到的“知识原语”(Knowledge primitive)何时能解决自主代理缺乏公司上下文的问题,从而真正替代部分人类协调工作?

Sources

Sign in / Sign up
ReadingResearchRadarInvestment framework
  • youtube.comYouTube: No, That's Not a Software Factory — Ryan Cooke, WorkOS↗
  • ai.engineerAI Engineer: Talk Transcript & Slides↗
  • workos.comWorkOS Blog: The self-driving codebase↗
  • workos.comWorkOS Blog: Agent Night demo recap↗
  • tubrief.ioTubrief: Video Summary↗

Coverage in this topic

  • WorkOS Engineer Ryan Cooke: Unlocking the Software Factory Is Not Infrastructure, But Encoding Engineering Workflows into Executable Automation2026-10-27News

    Product engineer Ryan Cooke reported at the 2026 AI Engineer World's Fair that an initial software factory setup combining a sandbox, model routing, and repository produced results nearly indistinguishable from engineers running coding tools locally. He argued the industry is optimizing for the wrong variables—PR count and AI code share measure production rate, not value delivery rate—and demonstrated how WorkOS encodes planning (Hilltop documents and PM agents), coordination (ticket dependencies and completion webhooks), and context (an internal MCP gateway) into automation, enabling agents to participate in the engineering workflow itself rather than merely generating code.