The awesome guys at DeepSeek recently released DeepSeek-Harness.
This is one great piece of software, for many reasons I won’t go into here, let’s just say it is VERY modular.
It use a well-thought-out and functional plugin system that allows you to switch almost every part of the harness without effort. It’s super hackable! And it’s nice.
I was planning to write a long blog post about the relevance of the high-level design of cognitive systems (I will soon, I’m forcing me to write here more often), this is a quick and extremely powerful way to experiment with those designs limited to coding (but that encompasses most of what a generalist system would look like).
I tried making some experiments with Pi and then OpenCode (you can find part of the config here). Using DSH feels so refreshing, I’m experimenting a bit: I was able to replicate quickly my OpenCode setup, without much of the limitation I found with it (regarding UI or agentic flow).
Now, I think there is one big mistake that DeekSeek made here: when they announced the project they stated clearly the intention to let users
experiment with it and share the plugins they make on GitHub with the topic dsh-plugin.
When I’m writing GitHub reports: “Here are 10,460 public repositories matching this topic… ”, which is great BUT you have to pay attention.
Those plugins are not sandboxed or checked in any measure and it’s EXTREMELY easy to get in the position where your agent suggests you to install a totally unknown repo with only some Chinese description that does “thing x you wanted in your harness”. I’ve given a quick look to some of these, some fetch resources on their first run and they could very easily distribute malware in ways that are very hard from the GitHub repos beforehand.
So, have fun guys. But always pay attention to what you (or your agent) do.