AI InfraAI Infra
模型与 Agent 运行的地基：执行沙箱、控制面与执行面、训练与推理的调度、数据管线和评测平台。The ground models and agents run on: execution sandboxes, control and execution planes, training and inference scheduling, data pipelines, and evaluation platforms.
- 实践过Hands-on
- 当前状态Current status
- XI
- 造物之书里的实验Experiments in the book
模型能力最终都落在基础设施上。执行离数据和环境越近，Agent 越快、越稳，也越容易被观察。Model capability ultimately rests on infrastructure. The closer execution is to the data and the environment, the faster, steadier, and more observable an agent becomes.
在 Eazo 参与 V4 到 V5 升级：把 Agent 执行从后端 Worker 迁到 E2B 沙箱，拆分控制面与执行面，前端经 WebSocket 直连沙箱；也参与 Skills 驱动的多 Agent 体系。At Eazo I worked on the V4 to V5 upgrade: moving agent execution from a backend worker into E2B sandboxes, splitting the control and execution planes, and connecting the frontend to the sandbox directly over WebSocket. I also worked on a Skills-driven multi-agent system.
- 01
让前端直连沙箱执行，多轮任务能省下多少等待？When the frontend connects directly to the sandbox, how much waiting does a multi-turn task save?
- 02
控制面与执行面拆开后，失败恢复和权限边界怎样设计？Once the control plane and execution plane are split, how should failure recovery and permission boundaries be designed?
- 03
一个小团队的评测平台，最少需要哪些组件才能支撑每天的回归？What is the minimum set of components an evaluation platform needs to support daily regression for a small team?
- 04
训练数据管线里，哪一步最容易悄悄引入质量问题？In a training data pipeline, which step most easily introduces quality problems unnoticed?
现在还不能证明什么What cannot be proven yet
- 公司内的架构细节不能公开，只能用开源组件独立复现In-company architecture details cannot be published; reproductions rely on open-source components alone
- 缺少大规模训练集群的实践No hands-on experience with large-scale training clusters