DeepSeek open-sourced a CLI tool called Harness this week, decomposing the AI Agent (an autonomous program that invokes tools to complete tasks) runtime into three pluggable Lego-like layers—but the team has officially labeled it a developer preview, and interfaces may change at any time.

What this is

Think of an Agent as a "worker" that operates the computer and calls tools on its own; Harness is that worker's workbench. It's not simply a model wrapped around a terminal—it turns every component—model adapters (connectors that bring different LLMs on board), tool registries, session archives, sandbox (isolated secure runtime) policies, and the main loop driving multi-turn calls—into swappable plugins.

The architecture splits into three layers. The outermost is a composition layer of Profiles and Bundles that stack capabilities in sequence; the middle is the runtime backbone handling data flow for a single conversation; the innermost layer uses Session logs as the source of truth. The team calls this design "everything is a plugin"—even the main loop itself is a plugin, and can be replaced. To change capabilities, you don't edit core code—you attach plugins and events.

Industry view

Our read: this step moves DeepSeek from "selling models" to "selling foundations." Chinese LLM vendors used to compete on parameters and benchmark scores; now they're digging ecosystem moats—open-sourcing infrastructure so developers build on their foundations.

Supporters argue the design hits enterprise needs squarely: auditable and replaceable—the two points compliance teams in finance and healthcare care about most. The clean three-layer split also means plugging in new models or tools requires no changes to the core repo.

Critics raise at least three concerns. First, DeepSeek itself states interfaces will see breaking changes—what you learn today may be obsolete next quarter; can enterprise production environments bet on that? Second, the underlying Cordis framework originated from the QQ bot community Koishi, with limited production-scale validation—financial-grade stability is questionable. Third, while the architecture is elegant, the learning curve is steep; ordinary developers may not want to invest the effort.

Impact on regular people

  • For enterprise IT: worth having the tech team evaluate the approach, but don't bet production on it short-term—wait for the interface to stabilize across one or two versions.
  • For individual professionals: non-developers don't need to learn this now; developers can watch whether the Cordis plugin ecosystem takes off, but don't rush to fork.
  • For consumer markets: no short-term impact visible; long-term, if the ecosystem matures, more AI Agent products may run on DeepSeek models.