I’ll be honest — when I first saw DeepSeek Harness (DSH) trending on GitHub, I figured it was just another AI repo with a lot of hype.
Star counts on GitHub can mean a few different things. Sometimes it’s respect for the founder. Sometimes people just bookmark a repo for later and forget about it. And sometimes — maybe most often — it’s genuine appreciation for the results.
But with DSH? I think people were reacting to something else. Something weirder.
Think about it. DeepSeek already has V4-Flash, which performs ridiculously well at 300B+ parameters while staying cheap and landing a great cache hit rate. They’re in a strong position. If all they wanted was solid numbers, they could have taken the safe path: more post-training along the lines of Codex or Claude Code. That would have been fine. Respectable, even.
They didn’t.
Instead, DeepSeek built a harness with four different modes, and the contrast between them is wild.
If “harness” is a new word for you, the plainest translation is this: it’s the rig that sits around the model. The model is the engine. The harness is the steering, the pedals, and the dashboard. Same engine, very different driving experience depending on what you bolt around it.
The minimal mode that beats the fancy one
You’ve got Minimal Mode, which is bare bones — persistent Bash and str_replace_editor. That’s it. Two tools. And in some environments, DeepSeek-V4-Pro actually performs best in this stripped-down setup.
Wait, what?
That flips the whole assumption on its head. More tools equals better, right? That’s been the logic. But when the model itself is already strong, adding tools means more choices, more complex decisions, and more places to get it wrong. Sometimes less really is more.
The mode that builds its own tools
Then there’s Creative Mode, which goes all the way in the opposite direction.
In Creative Mode, the agent can create, modify, and uninstall its own plugins on the fly. It can invent a tool it needs, tweak it if it isn’t working, or throw it out entirely. Need a hammer? Build one. Hammer not cutting it? Change it, or scrap it and start over.
That’s not the usual setup. The old way: humans build the tools, models learn to use them. The new way: give the model a goal, let it work out what tools it needs, build them, use them, and discard them when the job is done.
Honestly? That’s the part that gets me. The hot-swappable nature of it all. Flexibility over structure. Freedom over predictability.
Here’s the thing I keep coming back to, though: a plugin the agent writes for itself runs with the agent’s permissions. Which quietly changes the security question from “did we review the tools?” to “do we trust the thing that made the tools?” That’s not a reason to avoid it. It’s just the part I’d want spelled out before I turned one loose on a repo I care about.
Because flexibility also means chaos. And chaos is the one thing large language models have historically been bad at handling. That’s exactly why most agents get trained inside structured environments like Codex or Claude Code. Stability matters. Predictability matters.
DeepSeek seems to be asking the opposite question: how much freedom should we give these models?
And the fact that V4-Pro behaves differently across modes tells me they’re probably still working that out themselves. This isn’t a polished product. It reads like an experiment.
Which is why I think the real story here isn’t the benchmark scores. Benchmarks are temporary. They change every week. The interesting part is the question underneath: when the model is already this capable, do we still need to dictate how it gets things done? Or should we hand over more control and let it find the path?
If agents eventually move past just using tools that humans prepared in advance — if they can create, modify, and discard tools based on the task in front of them — then we’re not building better assistants anymore. We’re building something that figures things out on its own.
That’s a different kind of leap.
I went into more detail on why the agent architecture matters here, if you want the longer version.
I’m not sure this approach wins out. There’s a reason structured agents are the default — they’re safer, more reliable, and easier to debug. But watching DeepSeek test the edge like this makes me wonder if we’ve been too conservative in how we think about agent design.
Oh, and I should mention — I haven’t actually run DSH myself yet. I’m going off what I’ve read and what people are saying. I could be wrong about parts of it. But the direction? That seems clear.
Anyway, I’ll keep an eye on where this goes. Might try it next week. We’ll see.
