We've noticed that DeepSeek's Agent framework dsh, open-sourced this week on GitHub, hides a thesis: the next battleground for Agent products (AI programs that autonomously execute tasks) isn't the model—it's the architecture. Can "local tools" and "cloud sandboxes" share a single codebase?

What this is

dsh breaks every Agent capability—reading files, executing Shell commands, searching—into three segments:
  • Definition: the interface contract, specifying what this capability can do.
  • Provider: the actor doing the actual work. Local mode calls your computer's disk directly; remote mode operates inside a cloud sandbox (an isolated remote runtime).
  • Consumer: the tool entry exposed to the AI model—calls only the interface, doesn't care who sits behind it.
Source-code comments point out: "The essence of the three-tier split isn't a design pattern—it's dependency inversion (letting the upper layer depend only on abstractions, not concrete implementations)." The Consumer only recognizes the interface; the Provider can be hot-swapped (switched at runtime)—uninstall the local Provider, install the remote E2B Provider, and the same conversation instantly shifts from "operating your computer" to "operating a cloud VM." Neither the model code nor the tool code moves.

Industry view

Supporters read this as Chinese big-model firms making a real framework-layer investment. Once the model-parameter arms race converges, whoever controls the Agent abstraction layer controls the next-generation developer entry point. dsh's five Seams (join points) cover files, Shell, search, conversation, and logs—aligned with the internal architectural thinking at Anthropic and OpenAI, but released open-source, betting on ecosystem positioning.

Dissent exists. A senior architect commented on X: "Pretty Seams don't equal production-ready." TypeScript declaration merging (syntax that lets types auto-infer) is friendly to frontend engineers but a hidden barrier for enterprise backend teams—once DeepSeek's in-house dependency-injection framework Cordis upgrades, migration costs are uncertain. Another concern: DeepSeek points the "sandbox" to US startup E2B, leaving Chinese enterprise compliance questions unresolved.

Impact on regular people

For enterprise IT: When picking an Agent product, it's worth one more question: "Can it be deployed locally?" If the underlying layer is a cloud sandbox architecture, every conversation passes through a third party.

For individual professionals: The Agent assistant you use today may be quietly switching its backend from local to cloud. When handling pay stubs and client lists, watch the vendor's outbound network rules.

For the consumer market: The "privacy war" in AI products is shifting from "will the model remember me" to "whose hard drive does the file land on." Local-first products will regain selling points.