DeepSeek Harness Introduction

AI models can already independently complete complex coding tasks. The real challenge is no longer making models smarter, buthow to organize their capabilities into a truly usable Agent in a stable, reliable, and observable way.。

In August 2026, DeepSeek open-sourcedDeepSeek Harness（dsh）— aEverything is a Pluginthe Agent framework, which provided the answer to that question:Agent = Model + HarnessModels are the soul of an Agent, and Harness gives the Agent the ability to understand the environment, use tools, and work continuously in real-world scenarios.

---

What is DeepSeek Harness?

Underlying it is an open-source plugin systemCordisdriven by it, whose design philosophy corresponds to the paper "A Programming Paradigm for Spatiotemporal Composability". At runtime, there is no privileged kernel that requires patching: every capability registration is a reversible side effect, automatically undone when the plugin is unloaded. Therefore, the way to extend dsh is to mount new plugins alongside other plugins.

---

Why is a Harness needed?

The industry has reached an increasingly consistent conclusion:The bottleneck is not model intelligence, but infrastructure.In a 1-million-line code experiment by the OpenAI team, all output over five months was produced by Agents, with engineers not writing a single line of code.

LangChain merely optimized the external driving environment (document structures, verification loops, tracing systems), and boosted coding Agents' score on Terminal Bench 2.0 from 52.8% to 66.5%, jumping from 30th to 5th globally—Not a single parameter of the underlying model is changed.

DeepSeek Harness is precisely the productization of this consensus: rather than pursuing a stronger model, it focuses on what the model needs to run —constraints, feedback, tools, memory, and observability.Everything is built as composable plugin infrastructure, so every team can harness AI in its own way.

---

Three leaps in the AI engineering paradigm

To understand why Harness matters, we first need to see clearly how we got here step by step:

Paradigm	Core problem	Optimization target	Interaction mode
Prompt engineering	How to speak clearly	Prompt wording, format, examples	Q&A
Context engineering	How to feed information to AI	Documents, code snippets, historical conversations	Information injection → Generation
Mastering engineering	How to make Agents work reliably	Constraints, feedback loops, control systems	Human steers, Agent executes

DeepSeek Harness is the engineering philosophy of harnessingComplete productization— an open-source implementation that fully realizes constraints, feedback, observability, and replaceability, so that steering engineering no longer remains mere methodology, but becomes out-of-the-box infrastructure.

---

Core architecture: everything is a plugin

The architecture of DeepSeek Harness can be explained in one diagram:The Cordis kernel is at the center, with all capabilities surrounding it as plugins, collaborating through services and events.

The key mechanism that realizes this architecture at runtime isProfile + Bundle: a running dsh is a plugin tree composed of layers stacked in sequence at startup — first each composition package is applied in the order listed by the profile, then the profile'scordis.patch.yml, home-level patch, and finally arbitrary--patchoverlay. Use one command to view the full configuration tree actually launched on the machine:

---

Four operating modes

DeepSeek Harness provides four running modes out of the box, covering the full spectrum from a complete coding Agent to minimalized benchmark tests:

Mode	Positioning	Capability composition
Standard mode	Feature-complete coding Agent	File editing, Shell, file and web retrieval, Skills, plans, goals, subagents, and workflows
PTC mode	Code composition tool calls	It has all capabilities of the standard mode, and exposes tools through the Code Mode SDK—letting the model combine multi-step operations with a single TypeScript program.
Minimalist mode	Minimized benchmark testing	Retains only persistent bash andstr_replace_editorTwo tools for evaluating models in minimal environments.
Creative mode	Custom Agent preset	Possesses all capabilities of the standard mode, and provides runtime checks, plugin experimentation, and preset creation guidance — combine to create your own new modes.

---

Core features

Feature 1: Every run leaves a trace

Feature 2: Multi-form usage, run anywhere

Feature three: open and controllable, no privileged kernel

Feature four: model-agnostic, plug-and-play

---

Quick start

After installing Node.js, start the Web UI with one command:

npx @deepseek-ai/dsh web

You can also install from source:

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Using the Web UI requires only three steps:

Steps	Operation	Description
1	Configuration model	Open Settings → Models, fill in your DeepSeek API key and save. Model routing is immediately available without a restart.
2	Select workspace	Add and select the project directory where dsh was started (the session input box is unavailable before selection)
3	Run task	Send something like "Summarize this repository" — the agent reads/writes files, runs commands, delegates to sub-agents; operations that exceed permission policies will first ask for your approval

---

Relationship between the Harness and related tools

DeepSeek Harness is not just another "library for writing a few agents", but ratherA complete layer above SDKs and frameworks, solving "how Agents can run reliably". Positioning differences from currently popular tools:

Project	Positioning	Relationship with DeepSeek Harness
Claude Code	Closed-source commercial coding assistant	Feature parity, but dsh is fully open-source, self-hostable, and has replaceable capabilities.
Hermes Agent	Self-Evolving Personal Agent (Nous Research)	It focuses on cross-session memory and skill accumulation; dsh focuses on plugin-based composition and full observability—the two concepts are complementary.
OpenClaw	Local-first message-based Agent	It focuses on multi-channel access and digital sovereignty; dsh provides Web UI / headless / SDK multi-form support.
LangGraph / AutoGen / CrewAI	Agent building framework (programming library).	Frameworks solve "how to build"; dsh is a complete Harness — "how to run stably + how to replace capabilities"

other extensions