Skills Are Obsolete: DeepSeek Harness Pushes Self‑Evolving Agents to a New Stage

DeepSeek Harness v0.1 developer preview was released under the MIT license and is powered by the Cordis meta‑framework. Its core principle is that everything—models, tools, skills, sessions, sandbox, and file system—can be expressed as plugins.

In earlier designs, skill, MCP and the agent loop formed a hierarchy: skill was the outermost layer, MCP sat deeper, and the loop was hard‑coded in the host process, creating a fixed boundary similar to adding a motor to a bike without changing the frame.

DeepSeek Harness collapses these layers so that skill, MCP, loop and scheduler now share the same abstraction level and are all interchangeable plugins. Four runtime modes operate on a single kernel, and swapping a plugin replaces the previously immutable loop.

This eliminates the need for an external “watcher” that previously managed self‑iteration. The Cordis component model allows nested components where a child’s lifecycle follows its parent, removing the separate overseer.

The author likens this internal evolution to a self‑climbing construction machine, where the framework itself becomes part of the load‑bearing structure rather than an external scaffold.

With this new layer, the competitive focus shifts from building better agents to providing a robust kernel that safely lets others plug in their own loops, scheduling, and scoring mechanisms; whoever establishes this foundation first could become the common base for future agent products.

Strategically, DeepSeek already possesses a strong V4 model line, so Harness is not merely a clone of Claude Code but a foundation that enables agents to safely modify their own execution—a prerequisite for AGI‑level self‑evolution.

The architecture introduces risk: as plugins proliferate, a single faulty plugin can cascade into a full process restart. The Cordis paper linked in the article demonstrates that the proposed safeguards are feasible.

Two unresolved questions remain: (1) how to guarantee safe modification and rollback when a plugin fails, and (2) how to evaluate which modifications are beneficial. Harness addresses the first but leaves the second open, as the evaluation mechanism and its “judge” are still undefined.

In summary, DeepSeek Harness transforms the framework from a simple application layer into an Agent OS, and its future impact depends on both the robustness of its plugin infrastructure and the development of reliable evaluation feedback loops.