DeepSeek Harness 深潜：过度工程，还是为自进化而建？ Zhenjia 从核心边界切入，对比 Pydantic… — Lakphy with FE&AI 见闻 — TG.ME

DeepSeek Harness 深潜：过度工程，还是为自进化而建？ Zhenjia 从核心边界切入，对比 Pydantic AI、Claude Agent SDK、Eve 与 DSH，并解释 DSH 为何借助 Cordis 把模型适配、工具、持久化、压缩乃至 agent loop 都做成可运行时替换的插件。文章认为，这种“Everything is Plugin”在今天明显过度工程，却是在为 Harness 自进化与即时生效预留结构；真正尚未解决的瓶颈，是低频、昂贵且不完全客观的评测信号。 The deepseek of DeepSeek Harness: Overengineering or built for self-evolution? Zhenjia compares the core boundaries of Pydantic AI, Claude Agent SDK, Eve, and DeepSeek Harness, then connects DSH's Cordis-based Everything-is-Plugin architecture to runtime self-evolution. The verdict: it is overengineered for most teams today, but intentionally so; the harder unsolved problem is trustworthy evaluation before self-generated harness changes go live. https://zhenjia.dev/posts/the-deepseek-of-deepseek-harness