Now

Building

• —最近把主要工程精力放在 Agent Harness 和 Benchmark，持续研究 Context Router、Tool Registry、Permission Gate、Verifier、Trace，以及怎样让评测结果可复现。
• —md2wechat-skill 已达到 3,617 Stars。下一步准备适配 DeepSeek Harness，做一个 dsh-plugin，测试公众号发布流程怎样进入插件化 Agent Runtime。
• —发布 Weekly Wisereads，把 Readwise 每周高亮最多的文章、视频、PDF 和电子书做成中文深度解读，并跑通发现、研究、质量门禁和原子发布。
• —开始认真骑行和拍 Vlog，也在探索一款偏习惯、养成和反馈的 Agent-native 骑行 iOS App。希望让 AI 带来的效率，慢慢转成更好的生活节奏。
• —Most of my engineering attention is now on Agent Harnesses and benchmarks: context routers, tool registries, permission gates, verifiers, traces, and reproducible evaluation.
• —md2wechat-skill has reached 3,617 GitHub stars. The next experiment is a dsh-plugin for DeepSeek Harness, testing how the WeChat publishing workflow fits into a plugin-based Agent runtime.
• —Shipped Weekly Wisereads, turning Readwise's most-highlighted weekly articles, videos, PDFs, and curated ebooks into Chinese deep-reading reports with discovery, research, quality gates, and atomic publishing.
• —Started cycling seriously and filming vlogs. I am also exploring an Agent-native cycling iOS app centered on habits, progression, and feedback, while turning AI-enabled efficiency into a better pace of life.

Writing

Writing about Agent Harnesses, benchmarks, Skill engineering, open-source retrospectives, and Weekly Wisereads. Cycling is now part of the content plan too: first rides, city routes, and the question of where the time saved by AI actually goes.

Community

Continuing to run an AI Builder community around Agent Harnesses, benchmarks, open-source projects, md2wechat, Codex workflows, and lessons from building a real content growth system. Join my ZhiShiXingQiu: 杰尼·AI 实战圈. Join

Learning

• —继续拆解 DeepSeek Harness、Codex 和其他 Agent Runtime，重点看 Context、Memory、Tool、Permission、Verifier 和 Trace 怎样协同。
• —最近做了很多 Benchmark，正在研究怎样评估 Agent 找证据、用工具、修正错误和验证结论的完整过程。
• —把骑行当成一个长期实验，观察习惯、即时反馈和低摩擦操作怎样帮助人持续行动，也为 Agent-native 骑行产品积累真实输入。
• —Studying how context, memory, tools, permissions, verifiers, and traces work together across DeepSeek Harness, Codex, and other Agent runtimes.
• —After running many benchmarks, I am focusing on how to evaluate evidence gathering, tool use, error repair, and verification across the full Agent process.
• —Treating cycling as a long-term experiment in habits, immediate feedback, and low-friction interaction, while gathering real input for an Agent-native cycling product.