DeepSeek Just Open-Sourced a Full Agent Framework — and It Changes the Math on Building AI Agents

DeepSeek just released Harness, a developer-preview agent framework, and the developer community is paying attention — 438 points and 200 comments on Hacker News in five hours. The reason isn't just another AI model release. It's a fundamentally different approach to building agents.

What Is DeepSeek Harness?

DeepSeek Harness is an open-source agent framework built on the principle that everything is a plugin. The model is the soul of an agent, but the harness — the infrastructure that lets an agent understand its environment, use tools, and keep working — is what makes it practical.

The framework is built on the Cordis kernel, which manages plugin mounting, unmounting, and dependencies. Every agent capability lives in a plugin: models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and even the UI. Developers can select, swap, or extend any capability through configuration without touching the Harness source code.

This is a departure from how most agent frameworks work today. Frameworks like LangChain, CrewAI, or AutoGen bundle capabilities into monolithic packages. Want to swap the model? You might need to rewrite your tool definitions. Want a different sandbox? Good luck. DeepSeek Harness treats each of these as independent, composable plugins.

Four Runtime Modes

Harness ships with four runtime modes that cover the full spectrum of agent development:

Standard Mode

The full coding agent experience: file editing, shell access, file and web search, skills, planning, goals, subagents, and workflows. This is what you'd use for building production agents that need to accomplish complex multi-step tasks.

Code Mode

All Standard mode capabilities, but with tools exposed through a Code Mode SDK. Instead of making individual tool calls, the model writes TypeScript programs that orchestrate multiple rounds of tool calls in a single step. This is significant — it means the agent can express complex logic as code rather than as a sequence of natural language tool invocations.

Minimal Mode

Just two tools: a persistent bash shell and a file editor. This is designed for benchmarking models in a minimal environment, stripping away the infrastructure to measure raw model capability. If you're evaluating whether a new model can actually code, this is the mode that gives you a clean signal.

Creator Mode

Built for building itself. Creator mode lets you inspect the current runtime, test Cordis plugins in memory, and combine them into new modes. It's a meta-mode for framework developers who want to create custom agent presets.

Every Run Is Traceable

One of the most compelling features for production use is the session log. Everything the model sees is recorded in an append-only log: system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection.

The Trajectory view lets you inspect these records by source. Resume, fork, search, and replay all operate on the same event stream. This means you can:

• Debug failed agent runs by replaying the exact sequence of events
• Fork from any point to try a different approach without losing the original
• Search across sessions to find when and why a decision was made
• Audit agent behavior for compliance and safety review

For anyone building agents that make real-world decisions, this level of traceability is not a nice-to-have. It's a requirement. Most agent frameworks treat logging as an afterthought. DeepSeek Harness treats it as a core architectural principle.

Getting Started

The quick start is about as simple as it gets:

$ npx @deepseek-ai/dsh web

That launches the web UI. For a full source install:

$ git clone https://github.com/deepseek-ai/deepseek-harness

The project is MIT licensed and open source. Community plugins are already appearing at github.com/topics/dsh-plugin.

Why This Matters

The agent framework space is crowded. LangChain has mindshare. CrewAI has simplicity. AutoGen has Microsoft's backing. But DeepSeek Harness introduces something different: a genuinely composable architecture where you don't have to buy into a specific philosophy of how agents should work.

Want to use a DeepSeek model with a custom sandbox and your own tool definitions? Configure it. Want to swap in a GPT model with LangChain tools? There's probably a plugin for that. The Cordis kernel doesn't care what plugins you mount — it just manages the dependencies and the event flow.

The timing is also notable. DeepSeek Harness drops on the same day as Gemini 3.7 Flash, Google's new agent-optimized model. The AI industry is pivoting from "whose model is best" to "whose agent infrastructure is best." DeepSeek's answer is to make the infrastructure open, composable, and free.

The Plugin Ecosystem Play

DeepSeek is explicitly courting a plugin ecosystem. The dsh-plugin GitHub topic is already active, and the framework's MIT license means there's no commercial restriction on building and selling plugins. If you're a developer looking for an income opportunity, building specialized plugins for DeepSeek Harness — say, a plugin for a specific API, a custom sandbox implementation, or an industry-specific tool pack — could be a viable path.

This is the kind of platform play that creates new markets. When VS Code opened its extension API, it spawned an entire ecosystem of paid extensions. When WordPress introduced plugins, it created a multi-billion dollar marketplace. DeepSeek Harness is early, but the architecture is designed for exactly this kind of growth.

Limitations and Caveats

It's a developer preview. DeepSeek explicitly says "core plugins and APIs will continue to evolve." That's developer-speak for "things will break." If you're building production agents on Harness today, expect breaking changes and plan for migration.

The framework also assumes you're comfortable with TypeScript/JavaScript. The Code Mode SDK and plugin system are JavaScript-native. Python-first developers will need to adapt.

Finally, the model quality matters as much as the harness. Harness gives you the infrastructure, but you still need a capable model to drive it. The good news is that Harness is model-agnostic — you can plug in whatever model you want.

Bottom Line

DeepSeek Harness is the most thoughtfully designed agent framework I've seen in 2026. The everything-is-a-plugin architecture, combined with complete traceability and multiple runtime modes, addresses the real pain points that developers face when building production agents. Whether the plugin ecosystem takes off will determine its long-term success, but the foundation is solid.

If you're building AI agents and haven't looked at Harness yet, it's worth your time. At minimum, the traceability features alone will change how you debug agent behavior.