技术分享Technology Notes
从真实问题出发，记录 Agent、机器人、论文阅读与 AI 提效的设计、实现、评测和复盘。Practical notes on agents, robotics, paper reading and AI productivity, from problem definition through implementation and evaluation.
知识漫游Knowledge Roaming
从科学、历史和技术故事出发，追问现象背后的长期线索与真实证据。Following long-term questions and evidence through science, history and technology stories.
面经分享Career Notes
岗位拆解、项目追问、系统设计与复盘，仅通过密码访问。Role analysis, project deep dives, system design and interview retrospectives, available by password.
今日AI快讯Today's AI Brief
MLE（Machine Learning Engineering，机器学习工程）Agent 很少接受标签稀缺场景的评测MLE (Machine Learning Engineering) agents that are rarely evaluated when labels are scarce
02编码 Agent 无法判断哪些历史内容值得留给未知的下一任务coding agents being unable to decide which repository history is worth preserving for an unknown next task
03Gemini 4 Hydrogen，定位为日常 Agent 工作流的高性价比模型Gemini 4 Hydrogen as a cost-efficient model for everyday agent workflows
04Claude Duet 更像能整理上下文、提醒遗漏事项的协作伙伴Claude Duet acts more like a context-aware partner that catches missed items
05原生 Swift Foundation Models Agent 接口native Swift Foundation Models API for model-backed agents
最近更新Latest Notes
认真写少数值得留下的内容，而不是堆积更多信息。A few carefully written pieces, instead of more information for its own sake.
harness 解剖学Anatomy of a Harness
论文精读：十一个编码 harness 的源码级解剖——七个子系统、13 条跨系统观察、29 个复用模式，以及两条经受住复核的缺席结论：没有通用 agent 框架，也没有向量检索查代码。A deep read: a source-code anatomy of eleven coding harnesses, with seven subsystems, 13 cross-cutting observations, 29 patterns, and two absences that survived re-auditing.
阅读文章answer-me-with-htmlanswer-me-with-html
在 Codex 里安装 answer-me-with-html 并把它调成按需触发：两层防线（skill 描述 + 全局规则）、skills 路径不一致的坑，以及一套可复现的验证方法。Installing answer-me-with-html in Codex and tuning it to fire on demand: a two-layer guard, the skills-path mismatch, and a reproducible way to verify it.
阅读文章
自演化的共谋作弊Co-Cheating in Self-Evolution
论文精读：自演化搜索 agent 的共谋作弊——出题者与解题者逐渐在同一批错误上达成一致，内部奖励上升而外部正确率不动；CrossFit 用源级交叉拟合反馈把这条回流路径切断。A deep read: co-cheating in self-evolving search agents, where proposer and solver converge on shared errors, and CrossFit, which cuts the return path by cross-fitting feedback at the source level.
阅读文章想交流机器人、Agent 或论文？Interested in robotics, agents or research?