Hi, I'm Adong / 阿东玩 AI
Router Harness · Harness RSI (Recursive Self-Improvement)
Building products that are fun and useful.
?
BaiDu Senior R&D → Lenovo Agent Architect → ?
AI master's graduate from the University of Chinese Academy of Sciences · Institute of Automation; former BaiDu Senior R&D Engineer, now Lenovo Agent Architect / AI Native Innovation Lead.
My research focuses on Router Harness and Harness RSI (Recursive Self-Improvement): how agents route computation across models and execution environments, and how their harness can improve through execution feedback and evaluation.
Created AgentGuide 
Built a 1M+ technical audience across Douyin, WeChat and Xiaohongshu, delivered Agent talks to 300+ Lenovo colleagues, and received the Lenovo China AI Pioneer Award and an Outstanding performance rating.
Currently holding Agent Algorithm Expert opportunities from all three BAT companies.
中文介绍 / Chinese bio
中国科学院大学（中科院自动化所）人工智能硕士，曾任 BaiDu 高级研发工程师，现任 Lenovo Agent 架构师 / AI Native 创新组负责人。
主要研究 Router Harness 与 Harness 的递归自我改进（RSI）：面向长程 Agent 任务，研究模型选择、端云协同与执行状态管理，在质量、成本和时延约束下进行逐步路由；利用执行轨迹和评测反馈，探索 Harness 的编排逻辑、工具、记忆与 Skills 如何持续改进，并通过独立任务和可复现实验验证改进是否有效。
发起并主导 AgentGuide
持续在抖音、微信公众号和小红书分享技术实践，技术受众达 100 万+；面向 300+ 位联想同事开展 Agent 技术分享，获得 联想中国 AI 先锋奖与 Outstanding 绩效。
目前已获得 BAT 三家的 Agent 算法技术专家机会。
Selected work · Research & engineering · Community & impact · Homepage
- Router Harness — adaptive model selection and edge / cloud routing within the agent execution loop. I study how routing decisions interact with task state, context, tools and recovery over long horizons, balancing task quality, inference cost, latency and privacy.
- Harness RSI · Recursive Self-Improvement — using execution traces, failure analysis and evaluator feedback to improve the harness itself: orchestration, prompts, tools, memory, reusable skills and routing policies. I explore propose–test–select loops with fixed base models, versioned changes and rollback, measuring whether improvements transfer to unseen tasks.
Evaluation approach. Inspired by MLE-bench's evaluation of agents on machine learning engineering tasks, I prioritize task outcomes under explicit time and compute budgets, held-out evaluation, repeated runs, component ablations and reproducible traces. The aim is to measure whether routing and harness changes lead to reliable gains in quality and efficiency.
|
An end-to-end guide from Agent algorithms to runnable systems: Runtime, Agentic RL, RAG, data synthesis, tool use and evaluation. |
An implementation-first desktop Agent Harness covering Agent Loop, Tool, Memory, Context, Sidecar and sandbox auditing. |
- hybrid-router-oss — hybrid edge / cloud model routing with quality, cost, latency and privacy as first-class signals.
- agent-ready-template — a Harness-first template for coding agents to understand a repository, follow boundaries and take over development.
1M+ technical audience across multiple platforms. I share practical work on Agents, LLM systems, training, inference and evaluation through “Adong Plays with AI”.
An open community for LLMs, RAG, Agents, training, inference and evaluation — with learning paths, project notes, paper discussions and engineering exchange.
Join the community / read the guide
Delivered 4 Agent engineering talks to 300+ Lenovo colleagues and joined multiple AI hackathons as a speaker.
Outstanding · Lenovo China AI Pioneer Award · Annual Professional Contribution Project Award
More awards & activity records
- 2025 OpenAtom Developer Conference — Distinguished Contribution Award & invited speaker
- OpenAtom Open Source Community — Contributor of the Year
- CAS MCP Cloud Service Competition — Third Prize
- Algorithm Design & Programming Challenge — Silver Medal
Personal homepage · Open-source discussions · Contact / collaboration