One thing worth noting this week: DeepSeek has open-sourced its Agent framework as modular building blocks—signaling that the company is pivoting from "selling models" to "selling foundations."

What this is

The new product is called DeepSeek Harness (Agent refers to an AI that can autonomously plan and execute multi-step tasks; Harness can be understood as "the runtime skeleton for an Agent"). Released this week is the developer preview v0.1, with source code open-sourced under the MIT license.

The biggest highlight is the design principle: "everything is a plug-in." The capabilities an Agent needs—model invocation, tools, skills, sessions, sandbox, storage, scheduling, UI—are all broken into independently replaceable plug-ins, and developers can freely recombine them without modifying source code. The foundation is DeepSeek's in-house Cordis plug-in system, which only handles loading and dependency management.

It offers four runtime modes:

  • Standard mode: complete tool set
  • PTC mode (Programmatic Tool Calling): lets the model write a piece of code to combine multiple tools
  • Minimal mode: keeps only one shell and file editing tool, convenient for benchmark testing
  • Creative mode: allows in-memory experimentation and assembly of new modes at runtime

Startup requires just one command: npx @deepseek-ai/dsh web. The system prompts, chain-of-thought (the model's reasoning process), tool invocation results, and sub-Agent scheduling for each run are all written into an append-only log (a record-keeping method where entries can only be added, not modified), and can be replayed or branched in the Trajectory view.

Industry view

What's worth commending: this is scarce supply in the open-source Agent framework space. Anthropic's MCP and OpenAI's Function Calling are both moving toward "composability," but they're locked within their own ecosystems. DeepSeek opening up the foundation means enterprises can build custom Agents locally without being locked in.

But there are three reservations worth noting:

First, v0.1 is still a preview—the documentation is rough, stability is unverified, and the developer community is mostly watching and waiting.

Second, the cost of "everything is a plug-in" is "everything is a dependency." Once Cordis's upstream interface changes, all downstream plug-ins must be rewritten. Architectural flexibility may simultaneously become a long-term maintenance burden.

Third, open source doesn't equal a moat. Similar plug-in frameworks already exist in communities like LangChain and CrewAI. DeepSeek's real test is whether it can rally an active contributor ecosystem—and historically, this isn't something DeepSeek has excelled at.

Impact on regular people

For enterprise IT: no direct short-term benefit, but domestic AI infrastructure gains another piece of the puzzle—from "only being able to call model APIs" to "being able to build Agents locally," adding another option for compliance and data control.

For individual careers: it's still a developer toy for now; if the ecosystem takes shape, "knowing how to modify AI plug-ins" may become a basic skill like "knowing how to write Excel formulas."

For the consumer market: no impact felt yet. The real turning point will have to wait until someone builds a mass-market application on top of this framework.