DeepSeek Harness: The Agent Harness Where Everything Is a Plugin

DeepSeek released an agent harness with no core code. Every module, including the agent loop itself, is a plugin compiled at startup. The creator calls it the fastest growing GitHub repo in history, and the on-screen repo page backs the momentum: 115k stars and 11.2k forks on deepseek-harness shortly after release. The architectural bet is what matters. Instead of shipping a monolithic harness with extension points bolted on, DeepSeek shipped a plugin runtime called Cordis and built the entire harness out of plugins loaded in sequence.

No Core Code

The claim sounds like marketing until you see what it covers. In a typical harness, you can extend tools or swap models, but the loop that drives turns is fixed. In DeepSeek Harness, each of these is a plugin you can replace:

Module	What it controls
Model adapter	How the harness interfaces with the LLM directly
Tool registry	How the agent defines and registers tools
Session log	How the harness reads and writes transcripts
Agent loop	How the agent moves through turns

The final runtime you use is the composition of these plugins, combined and loaded at startup. There is no privileged harness code sitting underneath them.

Cordis, the Plugin Runtime

The video shows an analysis doc describing the harness as "a Cordis composition where capabilities register services, typed events, and reversible effects into a shared runtime context." Cordis is the small framework underneath: plugins mount into a shared context, expose services, and communicate through typed events. The official architecture docs state that every part of the product is a plugin, including the model adapter, tool registry, session log, and agent loop. That makes the plugin model load-bearing. It is how the harness exists, not just how users extend it.

The Session Log as Source of Truth

One frame shows the event model: step/start, user/message, assistant/chunk, tool/call, tool/result, step/end. Conversation history is a projection from the log rather than a separate data structure. System prompts, user messages, assistant chunks, tool calls, tool results, and injected contexts all become reconstructable facts. If you have ever fought a harness over how it truncates or rewrites transcripts, a swappable session-log plugin is the escape hatch.

How It Compares to Pi

The creator's reference point is the Pi agent, which follows a similar minimal-core philosophy. The difference: Pi still has core code, just a small amount. DeepSeek Harness removes the core entirely. If you want to modify how turns execute, you swap the agent loop plugin instead of forking harness internals.

Developer Preview, Not Production

The analysis doc on screen is direct about this in its "Risk and Caveats" section: this is not production software. It is a developer preview, architecturally special, and still unstable. The repo README says the same thing in capital letters: there will be compatibility-breaking changes. Treat it as a place to learn harness architecture and prototype, not as a foundation to ship on this month.

Key Takeaways

DeepSeek Harness builds the entire agent harness from plugins compiled at startup, with no core code
Cordis is the underlying runtime: plugins register services, typed events, and reversible effects into a shared context
The model adapter, tool registry, session log, and agent loop are all swappable
Session history is a projection from an event log, making transcripts reconstructable and the log plugin replaceable
Compared to Pi, which keeps a small core, DeepSeek Harness makes every layer modifiable
It is a developer preview with breaking changes expected, so it is for learning and prototyping right now

Resources

deepseek-ai/deepseek-harness -- The repo: "DeepSeek Harness: Everything is a Plugin"
DeepSeek Harness product page -- Official developer preview site
Cordis tutorial -- Docs for the plugin framework underneath the harness
awesome-deepseek-harness -- Community-curated plugins and ecosystem tools