8 plugins, under 200 lines of TypeScript code—this Juejin tutorial replicates the core architecture of DeepSeek Harness (an open-source Agent runtime framework), and our judgment is that the moat in the Agent infrastructure layer is already leaking.
The tutorial author uses Cordis (a minimalist open-source dependency injection framework, similar to a lighter Koa plugin mechanism) as the base, hooks in DeepSeek's conversation API, and assembles three building blocks—bash, fetch, and file search—running through the Agent loop where "the model decides which tool to call next." He openly acknowledges the gap is only in engineering polish—the skeleton itself is now open-sourced.
What this is
In essence, this tutorial is a "DIY Agent framework" manual. What it demonstrates is not a new product, but a structural judgment: when an Agent framework can be broken down into four parts—plugins + services + events + loops—the assembly threshold drops to a level any developer who knows TypeScript can tackle. In other words, the Agent framework is no longer a black box, but Lego.
Industry view
Optimists say this is good—more open infrastructure means more application-layer innovation. But the counterarguments are just as solid.
First, "runnable" is far from "production-ready"—error handling, concurrency safety, observability, and permission boundaries are all blank in the tutorial. The complexity of a production-grade framework is obscured by this diagram.
Second, the tutorial looks "simple" because it chose Cordis, a minimalist base; swap in LangChain or Semantic Kernel, and a few hundred lines wouldn't be enough. Equating "lightweight framework" with "low barrier" is a survivorship bias (the statistical illusion of seeing only successful cases while ignoring failed samples).
Third, and most worth watching—when the Agent framework itself is no longer scarce, value gets squeezed toward both ends: up to the underlying large models, down to vertical-scenario applications. The "Agent platform companies" in the middle will be caught in a pincer attack. That is the real industry problem.
Impact on regular people
For enterprise IT: companies wanting to build internal Agent tools have indeed seen a lower development barrier, but they need to separate "demo runs" from "production deployment"—there's a great deal of engineering work in between.
For individual careers: developers who can "assemble Agent frameworks" are worth more than those who only "call APIs." For non-technical roles, the signal is—the AI application layer is where the opportunity lies; don't fight over the model layer.
For consumer markets: C-end users won't feel this in the short term, but the indirect signal is that in the next 6 months, a wave of "AI assistant" products with similar underlying structures will flood the market—choose by scenario, not by technical jargon.