Deepseek Harness Preview link

DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI.It uses an architecture where everything is a plugin, and is powered by Cordis, whose design is described in A Programming Paradigm for Spatiotemporal Composability.

Developer preview:

DeepSeek Harness is currently in developer preview and is iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.

I’ve been running deepseek harness all day instead of claude code. Before using deepseek harness, it didnt sound compelling at all but it’s got some interesting ideas already. Love the visibility in to what an agent is doing. The seamless ability to toggle back-and-forth on read-only and write is nice. I could also see building up a bunch of agent profiles as well for specific needs like triaging prod, project specific things, etc. Finally, paired with ds4 flash on dgx spark, it was a pretty good working environment and would have been the best six months ago.

Back to my normal coding harnesses tomorrow though. I work in the cli most of my day, so a web app doesn’t fit for me. But look forward to seeing where this harness goes.

I used this quite extensively over the past 48 hours and I think this will end up being a great platform to build off of - however, there are just too many bugs that need to be addressed for me to start using this as a daily coding platform. The “everything is a plugin” is a great way to DIY an experience customized exactly for the user… sadly, many of the plugins I tried were buggy or flat out broken. This is def something to stay tuned to!

I spent a full day trying to set this up, and in the end I couldn’t get even half of what Hermes Desktop can do out of the box.

Plugins are still quite buggy, API keys are entered but sometimes not detected, generation speed is displayed poorly, and there are various other rough edges.

That said, I really like the idea behind the project. It still feels quite raw right now, but there is definitely something promising here. I can easily imagine it becoming as widely used as Hermes Desktop in a few months if development continues at this pace.

One thing I especially like is how clearly you can inspect the agent’s history: what it did, what response it received, and how it arrived at a particular conclusion. This makes debugging much easier and helps you quickly understand where something went wrong.

So overall, I’m looking forward to seeing how it develops. But in its current state, I personally wouldn’t recommend spending too much time trying to use it seriously yet.