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 it for later, and sometimes—maybe most often—it’s appreciation for the performance.
But with DSH? I think people were reacting to something else. Something weirder.
I mean, think about it. DeepSeek already has V4-Flash, which performs ridiculously well at 300B+ parameters while staying cheap and having a great cache hit rate. They’re in a strong position. If they just wanted solid results, they could’ve taken the safe path—more post-training along the lines of Codex or Claude Code. That would’ve been fine. Respectable, even.
But they didn’t.
Instead, DeepSeek did something that caught me off guard: they built a harness with four different modes. And the contrast between them is wild.
You’ve got Minimal Mode—which is bare bones. Just persistent Bash and str_replace_editor. That’s it. Two tools. And yet, in some environments, DeepSeek-V4-Pro actually performs best in this stripped-down mode.
Wait, what?
That flips the whole assumption on its head. More tools = better, right? That’s been the logic. But when the model itself is already strong, adding more tools means more choices, more complex decisions, and more room for errors. Sometimes less really is more.
And 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 if it needs one, tweak it if it’s not working, or toss it entirely. Need a hammer? Build one. Hammer not cutting it? Modify it, or scrap it and start over.
This isn’t your traditional agent setup. The old way: humans build the tools, models learn to use them. The new way: give the model a goal, let it figure out what tools it needs, build them, use them, and discard them when they’re done.
Honestly? That’s the part that gets me. It’s the hot-swappable nature of it all. Flexibility over structure. Freedom over predictability.
But here’s the thing—flexibility also means chaos. And chaos is the one thing large language models are historically bad at handling. That’s why most agents are trained in 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 performs differently across modes tells me they’re probably still figuring it out themselves. This isn’t a polished product. It feels like an experiment.
Which is why I think the real story here isn’t the benchmark scores. Benchmarks are temporary. They change every week. What’s actually interesting is the underlying question DSH is asking:
When the model is already this capable, do we really need to dictate how it gets things done? Or should we hand over more control and let it figure out the path?
If agents eventually move beyond just using tools that humans prepared in advance—if they can create, modify, and discard tools dynamically based on the task—then we’re not just building better assistants anymore. We’re building something that can figure things out on its own.
That’s a different kind of leap.
I’m not 100% sure this approach will win 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. Could be totally 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.
