DeepSeek dropped something new just yesterday — August 13. Not a model this time, but an agent harness called DeepSeek Harness. They open-sourced it under MIT and put it up on GitHub (https://github.com/deepseek-ai/deepseek-harness). I spun it up last night and spent a few hours with it. Here’s what I found.
Wait, I should back up. What’s an agent harness? It’s basically the thing that sits between a language model and the tools it needs to get stuff done — files, terminals, web searches, other agents. Think of it like the wiring that lets an AI actually do things instead of just talking about doing things. DeepSeek’s version is called dsh for short, and you can run it with one command if you have Node.js installed.
The big selling point? Everything is a plugin. Not some things. Everything. Models, tools, skills, sessions, sandboxes, storage, the agent loop itself — all plugins. You can swap any part out without touching the core code. That’s unusual. Most agent frameworks have a fixed core with extension points bolted on. This one has no privileged core at all. If you don’t like how the agent loop works, you replace it. If you want a different sandbox, you replace that too. The whole thing runs on Cordis, which is a plugin framework built around something called spatiotemporal composability. I won’t pretend I fully understand that paper, but the practical effect is clean: plugins load, plugins unload, nothing leaks.
DeepSeek Harness ships with four built-in modes. Standard mode is the full toolset — file editing, shell, web search, planning, subagents, the works. PTC mode does programmatic tool calling, which means the model writes TypeScript code to orchestrate multiple tool calls instead of making them one by one. That cuts down on round trips and feels more reliable. Minimal mode gives you just a shell and a file editor — useful for benchmarking. Creative mode lets you inspect the runtime and experiment with plugins in memory, then compose new presets on the fly. I spent most of my time in Standard mode because that’s what I’d actually use day-to-day.
Getting started is almost too easy. You run npx @deepseek-ai/dsh web and it spins up a web UI at http://127.0.0.1:3080. Then you add an API key in Settings → Models and pick a workspace folder. That’s it. No database setup, no config files, no yak shaving. The web UI is clean — not fancy, but functional. It shows you the session, the tools available, and the trajectory of what the agent did. I like that part. You can see exactly what the model saw at each step — system prompts, tool calls, results, context injections. Everything is append-only, so nothing gets silently overwritten. If something goes wrong, you can trace it back.
I threw a few tasks at it. The first one was simple: “Read the README in this project and tell me what it does.” It found the file, read it, and gave me a summary. Nothing impressive yet. Then I asked it to refactor a messy Python script I had lying around. It read the file, identified the problem, wrote a cleaner version, and ran the tests to verify. The tests failed on the first try. It read the error, fixed the issue, and ran them again. Passed on the second try. That’s the kind of thing that actually saves time — not generating code from scratch, but iterating on existing code without me babysitting every step.
I also tried the creative mode. Honestly? I’m still figuring it out. It lets you inspect the running runtime and experiment with plugins in memory. You can compose new presets without restarting. I’m not a plugin developer, but I can see the appeal if you’re building custom agent workflows. The documentation suggests spending an hour in creative mode just experimenting. I spent about twenty minutes and felt like I’d barely scratched the surface.
One thing worth noting — DeepSeek Harness is not a polished product yet. The team is very upfront about this. It’s in developer preview, they’re iterating fast, and there will be breaking changes. If you’re building something for production, pin your versions and watch the release notes. Don’t assume today’s config will work next month.
The bigger picture here is interesting. DeepSeek is moving beyond just selling models. With Harness, they’re offering an open framework for building agent systems — a competitor to things like Claude Code and OpenAI’s Codex. And they’re doing it under MIT, which means anyone can use it, fork it, or build on it without paying a license fee. That’s a smart play. The model itself is just one piece. The harness determines how that model interacts with the world. Control the harness, and you control a big part of the developer experience.
Oh, I should mention — DeepSeek also announced V4-Pro around the same time. It’s their flagship model with native support for agent workloads. The API pricing is changing too, with peak and off-peak rates kicking in soon. V4-Flash runs about $0.14 per million input tokens and $0.28 per million output. V4-Pro is higher — $0.435 for input, $0.87 for output. That’s not cheap, but it’s not insane either. If you’re just experimenting, the Flash tier is probably fine.
I’m not 100% sure I’d use DeepSeek Harness for serious work right now. The breaking changes warning gives me pause. But for experimenting, benchmarking, and plugin development? It’s genuinely worth a look. The plugin architecture is unusual and the trajectory logging is better than most alternatives I’ve seen. I’ll keep using it and see how it evolves.
If you’re building agents and you’re tired of frameworks that lock you into a single way of doing things, give it a shot. Run npx @deepseek-ai/dsh web, spend an hour in creative mode, and see what you think. Just don’t expect it to be stable yet. It’s not. But it’s heading somewhere interesting.