DeepSeek's open-sourced Agent framework Harness this week ships with 159 default plugins — a radical design where even the 'scheduler brain' itself is a plugin. It's a direct challenge to the 'fixed-core' approach of frameworks like LangChain and AutoGen. This week we noticed that, unlike Western vendors still racing on parameters, DeepSeek is betting on plug-and-play assemblability.
What this is
Agent frameworks (development platforms that let AI autonomously call tools and complete tasks) currently follow a mainstream pattern: a fixed core, with tools and model adapters bolted on. DeepSeek's Harness (DSH) chose a different path — model adapters are plugins, tool registries are plugins, session logs are plugins, and even the 'agent loop' that drives each turn of conversation is a plugin. There are no privileged components.
Underpinning this design is the Cordis engine, developed by the author of the Koishi framework and battle-tested by 4,000+ community plugins. DSH runs its 159-plugin 'mega-pack' on top of Cordis, letting developers swap any module without modifying source code.
Industry view
Supporters see this as a new trend: Chinese large-model companies are shifting from 'racing on parameter scale' to 'racing on ecosystem.' When you can freely assemble at the framework layer, even mainstream Western approaches start to look clunky — open source plus plug-in architecture becomes a viable path for overtaking on a curve.
Critics raise two concerns. First, 159 default plugins are an 'onboarding disaster' for newcomers — the abstraction layers sit too high, and the learning curve is steep. Second, the maintenance cadence of Cordis upstream directly determines the stability of DSH's entire architecture. Putting your lifeline on an engine you don't own may not be a safe bet.
Impact on regular people
- For enterprise IT: There's now an extra dimension to weigh when choosing an Agent framework — 'plugin-friendliness.' Future customization of in-house AI assistants may not require building from scratch; it'll mean assembling ready-made modules.
- For individual professionals: This doesn't directly affect your day job, but the AI customer service and intelligent assistants you use are likely built on this kind of architecture underneath — which means more room for tuning and extension.
- For the consumer market: A mature open-source ecosystem means lower costs for customizing AI services. Small and mid-sized companies can now access more flexible AI assistants, instead of being forced to buy the big vendors' all-in-one suites.