DeepSeek splits Agent into **"Evolvable Machine"**: Cordis paves the way for recursive self-improvement

According to Insight Beating monitoring, DeepSeek Harness has open-sourced, and the underlying framework Cordis has also released a design paper. This paper explains Harness's most radical and intriguing design: "Everything is a plugin." Model adapters, tool registration, Session Log, and even the driver for recursively calling tools in the Agent Loop are all just replaceable components. Harness does not have a lofty fixed core, and even the Agent Loop itself is just a replaceable plugin.

The real challenge is not breaking the code into plugins, but allowing these components to be continuously attached and detached at runtime without leaving behind uncontrolled dependencies and resources. Cordis divides the problem into "temporal composability" and "spatial composability." The former tracks the side effects generated by components, allowing resources registered by the plugin to be recycled when the plugin is unloaded; the latter manages dependencies between components and automatically adjusts their lifecycle when dependencies change. In short, the system needs to know not only what a component "does" but also what it "needs."

This design is directly related to recursive self-improvement. AI wanting to continuously improve itself doesn't necessarily have to start by tweaking the most challenging model weights. Cues, context, tools, workflows, and even the Agent Loop can themselves be the targets of improvement. Cordis first transformed Harness from a hardcoded program into a set of components that can be continuously combined, replaced, and unloaded.

DeepSeek has even entrusted the operation of this runtime interface to the Agent. The creation mode of Harness can inspect the current Cordis environment and dynamically define, load, and unload new components at runtime. Therefore, the Agent is not only calling existing tools; it can already directly transform the Harness it is in.

So, the most noteworthy aspect of DeepSeek Harness is not that it has created another Claude Code or Codex. It first turned the Agent into a machine suitable for continuous transformation.

If in the future, recursive self-improvement starts with Harness instead of model weights, Cordis provides a key foundation: allowing models, tools, contexts, Sessions, and even the Agent Loop to transition from fixed designs to machine-reconfigurable objects.