本文由 AI 分析生成
建立時間： 2026-05-25 來源： https://x.com/GenAI_is_real/status/2036266930290696599
Summary
A practitioner’s skeptical take on “harness engineering” as a newly coined term for what is fundamentally established software engineering applied to LLM work environments. The author built a multi-agent Q&A system for the SGLang community (how-to-sglang) and independently discovered the same patterns the harness engineering community names (progressive disclosure, repository as source of truth, structured routing), concluding that qualitative breakthroughs came from environment improvements, not stronger models — but that this doesn’t require new vocabulary.
一位實踐者對「框架工程」這個新術語的懷疑視角：它本質上是將已有的軟體工程原則應用於 LLM 工作環境。作者為 SGLang 社群建構了多代理問答系統，獨立發現了與框架工程社群相同的模式，得出突破來自環境改善而非更強模型的結論——但這不需要新詞彙。
Key Points
- Built multi-layered sub-domain expert architecture for SGLang: Expert Debating Manager decomposes questions, routes to specialized domain experts (advanced features, platforms, models/cookbooks), solves in parallel, synthesizes
- Stuffing all docs into one omniscient agent doesn’t work: attention scatters, quality degrades with context size
- Three independently-discovered patterns: progressive disclosure (each expert loads only its domain), repository as source of truth (all knowledge in markdown files in repo, no verbal agreements), structured routing (explicit table mapping question types to agents)
- Critical observation: every qualitative breakthrough on how-to-sglang came from environment improvements (better knowledge partitioning, routing, feedback loops), never from swapping a stronger model
- Open question: if model capabilities scale enough to build their own environments, will current design principles become irrelevant?
Insights
The observation that “same software engineering principles, different context” is genuinely useful pushback against the tendency to rebrand established practices. The SGLang multi-agent architecture is a concrete, real production example of what harness engineering advocates describe in abstract. The open question about AI self-designing environments is the most interesting thread — OpenClaw going from 400k to 1M lines in a month driven by AI suggests this transition may be closer than expected.
Connections
Raw Excerpt
What I actually learned from how-to-sglang can be stated without any new vocabulary: Information fed to agents should be minimal and precise, not maximal. Complex systems should be split into specialized sub-modules, not built as omniscient agents.