来自 ICML 的态势感知：关于 Code Agent
A field note from ICML on why Code Agent became the dense common topic, and why Code-style execution loops may become a general form for next-generation agent capability.
复利的第三范式
A mini-book essay on capital compounding, skill compounding, and how AI tools may form a third compounding structure.
分工之后
A mini-book essay on how the Agent era challenges the economic intuition inherited from The Wealth of Nations.
手机端 Agent 驱动同一台手机的 GUI 操作闭环
A small experiment where a phone-side Codex session asks desktop Codex to operate the same Android phone GUI through USB ADB.
跨 Harness 的上下文压缩与召回
A lightweight context layer that compresses, recalls, and transfers task state across sessions, tools, and agent entrypoints.
面向 Agent 工作流的后台并行 GUI 控制
A local macOS execution layer that lets agents inspect and operate real desktop apps in the background without interrupting your work.
面向 Agent 工作流的语音编程 Harness
A self-improving voice-coding harness that turns spoken intent into structured agent instructions, feedback loops, and reusable traces.
AI-native Slides：Agent 时代的幻灯片制作最佳实践
A personal field note on AI-native slide making across PPTX, Slidev, Feishu Docs, Beamer, and image-first versus editable-first workflows.
Linear Attention 作为可泛化记忆机制
A blog-format public draft on linear attention as a modular memory mechanism for agent pipelines.