GitHub's plan for Agents — Kyle Daigle, GitHub
English summary
GitHub COO Kyle Daigle discusses how AI agents are transforming software development, leading to a 1400% increase in code shipped by agents by 2026. This growth is straining GitHub's infrastructure, designed for human-paced development, causing reliability issues. Daigle explains GitHub's internal AI workflows, the shift from mega-skills to micro-skills, and the evolution of Copilot from code completion to a full agent platform. He also addresses challenges in open source trust, pull request verification, and the need for ambient AI that understands broader context.
Chinese summary
GitHub 首席运营官 Kyle Daigle 讨论了 AI 智能体如何改变软件开发，预计到 2026 年智能体推送的代码量将增长 1400%。这种增长对原本为人类开发者设计的 GitHub 基础设施造成了压力，导致可靠性问题。Daigle 介绍了 GitHub 内部的 AI 工作流、从大型技能到微技能的转变，以及 Copilot 从代码补全到完整智能体平台的演进。他还探讨了开源信任、拉取请求验证以及需要理解更广泛上下文的 ambient AI 等挑战。
Key points
GitHub is experiencing 14x commit growth due to AI agents, straining infrastructure and causing outages.
由于 AI 智能体，GitHub 的提交量增长了 14 倍，给基础设施带来压力并导致停机。
GitHub shifted from fine-tuning Copilot models to building a full coding agent platform with CLI, desktop app, and cloud agents.
GitHub 从微调 Copilot 模型转向构建完整的编码智能体平台，包括 CLI、桌面应用和云端智能体。
Daigle advocates for micro-skills over mega-skills to avoid brittle AI workflows and enable easier customization.
Daigle 提倡使用微技能而非大型技能，以避免脆弱的 AI 工作流，并便于定制。
Open source trust is challenged by AI-generated slop PRs; GitHub explores vouching systems and social signals.
AI 生成的垃圾拉取请求挑战了开源信任；GitHub 正在探索担保系统和社会信号。
Ambient AI and tools like OpenClaw are seen as the future, needing OS-level sandboxing and broader context awareness.
环境 AI 和 OpenClaw 等工具被视为未来方向，需要操作系统级沙箱和更广泛的上下文感知。