**[DeepSeek Breaks Down Agent into an "Evolvable Machine": Cordis Paves the Way for Recursive Self-Improvement]** According to monitoring by 动察 Beating, alongside the open-sourcing of DeepSeek Harness, the underlying framework Cordis has also released its design paper. This paper explains Harness's most radical and compelling design: "Everything is a plugin." Model adapters, tool registration, session logs, and even the Agent Loop that drives the model to repeatedly invoke tools are all replaceable components. Harness does not have a fixed, overarching core—even the Agent Loop itself is just a replaceable plugin. The real challenge is not breaking the code into plugins but enabling these components to be continuously installed and uninstalled during runtime without leaving behind uncontrollable dependencies or resource leaks. Cordis breaks the problem into "temporal composability" and "spatial composability." The former tracks the side effects generated by components, ensuring resources registered by a plugin are reclaimed when it is uninstalled. The latter manages dependencies between components, automatically adjusting lifecycles when dependencies change. In short, the system not only knows "what a component does" but also "what it needs." This design is directly tied to recursive self-improvement. For AI to continuously improve itself, it doesn't necessarily need to start with the most challenging aspect—modifying model weights. Prompts, contexts, tools, workflows, and even the Agent Loop itself can all become targets for improvement. Cordis first transforms Harness from a rigid program into a set of components that can be continuously recombined, replaced, and uninstalled. DeepSeek has even handed over the interface for operating this runtime system to the Agent. Harness's creation mode can inspect the current Cordis environment and temporarily define, load, and unload new components during runtime. As a result, the Agent is no longer merely invoking existing tools—it can now directly modify the Harness it operates within. This is why the most noteworthy aspect of DeepSeek Harness is not that it has created another Claude Code or Codex. Instead, it has turned the Agent into a machine designed for continuous transformation. If recursive self-improvement in the future starts with Harness rather than model weights, Cordis provides a critical foundation: enabling models, tools, contexts, sessions, and even Agent Loops to shift from fixed designs to objects that can be recomposed by machines. [Original Link]