Summary
Technical writer at yuanchaofa.com. Writes about AI agent engineering and LLM infrastructure in Chinese, covering harness engineering, inference optimization, and practical agent improvement strategies.
Key Points
-
Authored “Harness Engineering — Agent 不好用，也许不是模型的问题” — arguing performance issues are harness problems, not model problems
-
Proposed a three-level engineering hierarchy: Prompt Engineering → Context Engineering → Harness Engineering (expanding scope, not replacement)
-
Distinguished transient harness designs (compensate model limitations, will become obsolete) from persistent ones (physical constraints like storage, sandboxing, version control)
-
Key insight: harness execution trajectories become training data — co-evolution between environment and model
-
Also authored “理解 KV Cache 与 Prompt Caching” — explaining KV Cache, Prefill/Decode phases, and Prompt Caching with architectural implications for Agent systems
-
Connects inference-level concerns (Prompt Caching prefix matching) to Agent system architecture design
-
Also authored “RAG 进化之路：传统 RAG 到工具与强化学习双轮驱动的 Agentic RAG” — demystifying Agentic RAG through two implementation paths: tool-driven (Chatbox) and RL-driven (Search-R1)
Open Questions
None yet
Evidence Timeline
- 2026-04-07: First encountered via harness engineering article (published 2026-03-14)
- 2026-04-07: Second article ingested — KV Cache and Prompt Caching (published 2026-02-21). Expands scope from harness engineering to LLM inference fundamentals
- 2026-04-08: Fifth article ingested — RAG evolution from traditional to agentic (published 2025-10-03). Covers tool-driven and RL-driven Agentic RAG
相关页面
agentic-rag, chaoyue-he, gleb-rodionov, jinse-chuanshuo-dacongrming, agentic-rag-as-harness