DeepSeek Harness: everything is a plugin, and that's the whole point
DeepSeek just open-sourced its own agent harness under MIT, and 38,000 people starred it before most of us finished reading the README. It's called dsh, it's TypeScript, and it lands one day after V4 Pro went GA. The pitch is four words: everything is a plugin.
Not marketing-speak. Models, tools, skills, sessions, sandboxes, storage, the agent loop itself, scheduling, and the UI are all plugins you swap in configuration without touching a line of the harness source. It sits on Cordis, a composable kernel that handles plugin dependencies and lets plugins talk to each other through services and events. Four runtime modes ship with it. Standard is the full toolset. Code mode has the model write TypeScript programs that orchestrate many rounds of tool calls instead of emitting one call at a time. Minimal strips down to bash plus a file editor, which exists so you can benchmark a model without the harness flattering it. Creator lets you build your own presets with runtime inspection.
The part nobody is talking about enough is the session log. Everything the model sees goes into an append-only event stream: system prompts, reasoning, tool calls and their results, subagent scheduling, and every single context injection, tagged by source. Resume, fork, search and replay all run on that same stream. Anyone who has tried to debug why an agent went sideways on turn 60 knows that context is currently a black box assembled by a dozen invisible hands. DeepSeek made it a log file you can read.
Read the intent here. A year of benchmark fights taught everyone that the harness matters as much as the model, and the good harnesses were closed. Claude Code, Codex, Cursor. DeepSeek's move is to give away the shell so that swapping the brain becomes a config line, which is exactly the world a model company with the cheapest tokens wants to live in. Minimal mode is the tell. They built a mode whose only job is to measure models fairly, and they ship the cheapest frontier model.
Developer preview, so expect breaking changes. npx @deepseek-ai/dsh web gets you a local web UI on port 3080. Code at github.com/deepseek-ai/deepseek-harness, docs at deepseek.com/harness.
← Back to all articles
Not marketing-speak. Models, tools, skills, sessions, sandboxes, storage, the agent loop itself, scheduling, and the UI are all plugins you swap in configuration without touching a line of the harness source. It sits on Cordis, a composable kernel that handles plugin dependencies and lets plugins talk to each other through services and events. Four runtime modes ship with it. Standard is the full toolset. Code mode has the model write TypeScript programs that orchestrate many rounds of tool calls instead of emitting one call at a time. Minimal strips down to bash plus a file editor, which exists so you can benchmark a model without the harness flattering it. Creator lets you build your own presets with runtime inspection.
The part nobody is talking about enough is the session log. Everything the model sees goes into an append-only event stream: system prompts, reasoning, tool calls and their results, subagent scheduling, and every single context injection, tagged by source. Resume, fork, search and replay all run on that same stream. Anyone who has tried to debug why an agent went sideways on turn 60 knows that context is currently a black box assembled by a dozen invisible hands. DeepSeek made it a log file you can read.
Read the intent here. A year of benchmark fights taught everyone that the harness matters as much as the model, and the good harnesses were closed. Claude Code, Codex, Cursor. DeepSeek's move is to give away the shell so that swapping the brain becomes a config line, which is exactly the world a model company with the cheapest tokens wants to live in. Minimal mode is the tell. They built a mode whose only job is to measure models fairly, and they ship the cheapest frontier model.
Developer preview, so expect breaking changes. npx @deepseek-ai/dsh web gets you a local web UI on port 3080. Code at github.com/deepseek-ai/deepseek-harness, docs at deepseek.com/harness.
Comments