DeepSeek's open-source Agent framework Harness (an Agent is an AI program that can autonomously execute multi-step tasks) is being dissected chapter by chapter in the Chinese tech community — one tutorial on Juejin (掘金, a major Chinese developer platform) has already reached Chapter 10. It starts with the simplest "greeting" tool, then walks developers through registering tools, adding configs, attaching event listeners (HMR — Hot Module Replacement — code changes take effect without restarting), and packaging everything into a plugin bundle. What we find worth watching is the phenomenon itself: when a company's open-source framework can inspire external developers to spontaneously write this depth of tutorial series, we know the ecosystem flywheel is already spinning.
What this is
The tutorial's subject is Harness, the open-source project DeepSeek hosts on GitHub. dsh is its command-line entry point, running on top of Cordis (a microkernel-style service orchestration framework) — every tool, config, and listener registers as an independent "Fiber", loaded according to dependency relationships. Writing a new tool plugin is essentially writing a TypeScript file, declaring parameters and return values via defineTool, then loading it through cordis.yml. The most critical engineering detail: the entire system supports HMR — change one line in a plugin and you see the effect immediately in the browser, no Agent process restart needed.
Industry view
One view holds that this represents DeepSeek's strategic upgrade: as the gap between models narrows, the real moat — a long-term advantage that's hard to replicate — is the developer ecosystem. Whoever gets more third parties to write custom tools and integrate with their Agent will lock in the next wave of enterprise AI applications — the logic mirrors how Apple built durable defensibility through the App Store. Another view is far more measured — open-source Agent frameworks are not scarce; LangChain (the dominant overseas AI application orchestration framework) and LlamaIndex (focused on retrieval-augmented generation) actually started earlier than dsh and have higher overseas recognition. Open-sourcing the model, framework, and plugin specs together will indeed rally developers in the short term, but how to convert ecosystem traffic into commercial revenue over the long term remains, as we see it, an unanswered question.
Impact on regular people
For enterprise IT: over the next 12-18 months, when evaluating Agent vendors, "does it offer a mature plugin ecosystem" will matter more than "how many points higher the model benchmarks score" — real enterprise scenarios always need custom tools, not generic chatbots. For individual careers: Agent development roles (writing tools, writing workflows) are moving from geek niche to legitimate technical profession — engineers who can write plugins and integrate APIs will become more valuable. For consumer markets: end users won't feel any change for now, but the underlying ecosystem battles will, in our view, determine what your AI assistant can and cannot do two to three years from now.