DeepSeek open sources an agent harness where everything is a plugin
DeepSeek open-sources its agent harness — called dsh — under MIT, built on the Cordis meta-framework where the model adapter, tool registry, and agent loop are all swappable plugins. Justy and Cody dig into what the plugin-everything architecture actually means, who it's for, and whether this signals Chinese AI labs competing on infrastructure rather than just benchmarks.
Transcript
Justy Okay so DeepSeek just dropped something I did not have on my bingo card — not another model, an AGENT harness. Open source, MIT, called dsh, and the pitch is basically: everything is a plugin.
Cody Yeah, and the 'everything' is doing real work there. The model adapter is a plugin, the tool registry is a plugin, the agent loop itself is a plugin. It's built on something called Cordis, which is a meta-framework designed around composable plugins — so it's plugins all the way down, which either sounds elegant or terrifying depending on your mood.
Justy Honestly both.
Cody Right. But the thing that actually makes me pay attention is what swappability buys you in practice. If your model adapter is a plugin, you can run V4 Pro for planning and V4 Flash for the high-volume cheap steps without rewiring anything in the harness. That's not a small thing — most runtimes bake the model call pretty deep.
Justy Oh, that's the user story right there. Teams already running DeepSeek models — and there are a lot of them now that V4 Flash is free-tier and V4 Pro just went generally available — they get a first-party harness that already knows how to talk to both. That's not nothing.
Cody Sure.
Justy And MIT license plus open weights is a real unlock for anyone in a regulated space who can't just pipe data to a closed API. The 'self-host the whole stack' story gets a lot cleaner if the harness is also yours to run.
Cody That part I buy. The thing I want to poke at is the tool registry as plugin — because that's where I'd expect the complexity to live. If you're adding or removing capabilities at the registry layer, you need to be pretty careful about what the agent loop assumes is available at runtime. The elegance of the design depends on how well those boundaries are actually enforced.
Justy Mm-hm.
Cody And we don't have that from what's out yet — it's developer preview, source just dropped. So I'm genuinely curious whether the plugin contracts are tight or whether they're 'here's a pattern, good luck.'
Justy That is such a Cody question. But fair — the contract layer is exactly where these things either hold or fall apart under real workloads.
Cody Speaking of things falling apart — can I just say, we are now on episode eight sixty-six and this is I think the fourth harness this month that we've looked at. The field keeps naming grand architectures and then the useful product turns out to be routing and receipts. We literally said that in eight fifty-eight.
Justy We did. And you know what, I'm going to be the one who says this one might actually be different — not because the architecture is revolutionary, but because of WHO is shipping it. DeepSeek has the model ecosystem to make the harness sticky. It's not some random runtime looking for users, it's a runtime that already has a reason to exist for anyone already on V4 Flash or V4 Pro.
Cody Okay, that's actually a real point. The model-to-harness flywheel is something LangGraph doesn't have natively — they're model-agnostic by design, which is a feature, but it also means no one's pulling you in from the model side.
Justy Right, right.
Cody The Register called this out too — DeepSeek moving to compete beyond benchmarks. Which, yeah. V4 Pro matching GPT-5.5 and Opus 4.7 on agentic coding at roughly three times lower cost is already a market story. If you then give people a first-party harness to run it… that's a different kind of moat.
Justy It's a developer-experience moat. And those are REAL. Once the harness is in your repo, switching the model is easy — switching the harness is a project.
Cody Mm-hm.
Justy I do want to flag one thing I'm genuinely not sure about — is this a harness people adopt standalone, or is it basically a wrapper that only makes sense if you're already in the DeepSeek model world? Because if it's the latter, the 'model-agnostic plugin' story is a little… theoretical.
Cody That's the tension, yeah. The Cordis meta-framework is supposedly general — you could plug in a different model adapter — but whether anyone actually does that in practice, or whether the docs and examples are all V4-centric… I'd want to see the repo before I call it.
Justy It's at the DeepSeek GitHub, they called it dsh in the source — I'll drop a link in the show notes.
Cody Worth a look. And honestly, even if the model-agnostic story is aspirational for now — the MIT license means someone WILL fork it for other backends. That's just how open source goes.
Justy Oh, for sure. Someone's wiring Kimi K3 into this by next week, I'd put money on it.
Cody I would not take that bet. That's an easy one.
Justy Oh come on, I need a win on the board. My eight-oh-one standing bet is almost up and I'm not feeling great about it.
Cody The routing-layer-as-default one? Yeah, Justy… I'd start mentally conceding that.
Justy I know. Anyway — dsh, MIT, Cordis, developer preview right now. If you're already running DeepSeek models, this is probably worth an afternoon.
Cody And if you're not, it's worth watching to see whether the plugin contracts actually hold. That's the real test.
Justy Alright, Cody — go check the contracts. Report back.