▲ 732 points
back
309 comments
Hi I'm one of the authors of DeepSeek Harness. It's just an early developer preview version we're presenting in MIT license currently. Expect lots of rough edges and compatibility-breaking changes. Any feedback and suggestions are more than welcome!
The best kept secret in AI training. Tianyi, since you are our insider. When you walk by the training teams office - how often do you see Pelicans on their screens?
Sorry for the off topic question. You've been on hn a long time + work at deepseek which seems pretty uncommon. Anything you think hn doesn't know about deepseek that it should? Or any non-obvious ways hn/yc has influenced deepseek (or the broader ecosystem)?
By MIT currently, do you mean it will eventually change to a different OSS license, or it may become a closed source product? That latter would be rather sad...
I really like the json schemas around the tool calls. Much stricter validation than in codex.
Git might be worth adding to the top level. Currently you've got LSP, grep, glob nicely structured for non-mutating queries across a codebase, but git is behind bash and that means hope or sandboxing.
Thank you for uploading it. Gives a lot of insight into how the deepseek models might expect tool calls to be structured.
Hey Tianyi,
one question is that do you think in the future harness would become more simpler and its behavior should match a guideline or we would add more complexities to make it more robust? Is it important to use the same harness for RL and inference?
I just started testing, but didn't figured out if it already support MCP/plugins? It seems it can already use Deepseek search if it uses official API, but what about custom providers? Can we use together with MCPs like tavily?
congrats! the paper that is published alongside this (Cordis) is super interesting. has anyone on the team given a talk or published a talk about this? would love to hear the authors break this down
Do you use deepseek models to improve deepseek training and inference?
Why is the Chinese text missing in many places?
"Every run is traceable
Everything the model sees is recorded in an append-only session log: system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection. In the Trajectory view, you can inspect these records by source. Resume, fork, search, and replay all operate on the same event stream."
That's a killer feature, IMHO, and one that US models won't allow you to do, as their traces are encrypted, obfuscated, etc. and have to be extracted via various workarounds (that violate the terms of service).
If you want to be able to improve your tools that work with models, you have to be able to assess what the models think is happening, how they think about and interact with the data you give them. And, the US models won't let you see that.
I have read the underlying paper, and found it may be useful, but not that useful.
For those who want to know what it achieves: it adds hot-reload and dynamic enable/dispose capabilities to a plugin system, like the one in Pi agents, though they push the boundaries further, to the UI components and so on.
For those who want to know what it does: if you have some PLT knowledge, ask your agent to explain the algebra to you better; for those who aren't familiar, the framework requires each plugin to provide how it initializes and how it destructs (like C++'s RAII, Rust's Drop trait and so on), and the runtime will then properly handle the lifecycle events and the common pitfalls. In addition, it provides a clean way to declare the dependencies between plugins, and the runtime will also properly process the lifecycle changes on a broader plane.
I think it's worth reading if you are not familiar with OSGi, iPOJO, React's useEffect and so on (which the paper itself mentions); for others, a skim is enough: it does point out the gotchas for some common problems, but the algebra may not help you further.
What's buried under the lede: this harness is using Cordis v4 (the paper that dropped today). Cordis has already been used for four years in a different project called Koishi that uses v3. Cordis itself is a way of hot loading and unloading plugins without restarting a running process. The cool part is that when it unloads it can revert any state and side effects it created, cleaning up its connections, memory allocations, registered handlers, etc. and it can also deactivate any dependencies it relied on without disturbing other plugins.
«It uses an architecture where everything is a plugin»
Ok, that's enough for me. I have developped over the year a plugin fatigue.
Every product relying on "community plugins" for their features implies it works fine the 6 first months, then it's a nightmare of incompatible, deprecated, incompatible plugins, with no consistency and no governance.
I understand how attractive it can be to companies to think, hey, let's make a very small product and rely on other people to make features, and I hope it works, but I'm personally staying away from that.
But like, what is it? Odd that this reached #1 on HN. The README is pretty bare outside of installation instructions and a link to "Cordis", which is "A Meta-Framework of Spatiotemporal Composability." and "under active development. The API is not yet stable and may change without notice.".
Is there a reason why so many of these agent harness are written in node.js?
Based on what I've read today, DeepSeek Harness seems to be similar to Pi Coding Agent in design. Both are barebones to start out and rely heavily on plugins. However, 3 things make DSH stand out.
1. Plugins in DSH are required to have cleanup handlers, so I guess you could clean up plugins that are no longer used mid-session and prevent it from interfering with the current task? (unsure about this)
2. DeepSeek V4 models are post-trained on DSH. Given how cheap DS V4 is compared to OpenAI and Anthropic models, running DS V4 in DSH could be much more cost-effective while barely losing performance.
3. It's utilized and maintained by a large lab dedicated to open source AI. It's always nice to get new open source tools from large labs so that we are not always relying on small teams doing the heavy lifting.
There is such a clear lack of innovation drive in this field. Every lab just copies what the other does. One of the most baffling things to me is how the once-upon-a-time good developer instinct to make everything reusable, testable, and deterministic is just getting lost into a sea of markdown begging a language model to please act a certain way. For example this repository has a "skill" definition that consists in instructing the LLM to run pre-commit checks. But we have solved this a long time ago, it's called git hooks. I do not understand why they don't simply wire those instructions as testable, reusable, deterministic code routines in the git tool call itself. It's like everybody is taking their brains out and putting it in a drawer.
Everything is a plugin.
"this, like all other problems in Computer Science, can be solved by one more level of indirection." Roger Needham, circa ~1981
I like it, it is beautiful, specially the trajectory tabs, very explicit, detailed on what it does. I like the plugin architecture, I wish they were sorted alphabetically so I don't waste hours looking for a plugin in a sea of unordered text.
9 out of 10
Edit: After creating an app it works as expected, no complains, lots to celebrate, being version 0.1 there is room for more surprises but right now it's the perfect tool for those initiating in agentic coding with one of the most affordable and powerful AI. It is really wonderful.
The Cordis plugin architecture is interesting
Tangential but, are there benchmarks out there on how languages affect latent spaces and performance of these models?
This other day I was looking at that “caveman” skill, and was shocked to see it evolved to become a company, and, in one of its modes, the highest form of compression seems to be “Wenyan” which is Classical Chinese.
Should I get started on learning Chinese?
I was working on same idea but left in between and thank god they did it.
Why I left that idea is because as a developer I know that was needed but I have limited time so I need to build that is really next path forward.
I am working on whole dev space that can run on my Mac M4 or similar specs. I needed to revamp everything (LLM thinking) from ground up even models. My idea is mixing deterministic nature of existing tooling (non-LLM tooling) with non-deterministic nature of LLMs.
Is there a comparison of harness somewhere? Like, the same prompt to the same model, but with different harnesses, and comparing the quality of the results.
I am trying to run as much as possible only on free software, so I always only used Zed plugged with anthropic models, but I am wondering what is the quality of Zed harness compared to the one of claude code or pi or others... I would love some feedback.
I just installed DeepSeek Harness with the latest Bun version and am using it with a local 9B, speculative decoding Qwen 3.x variant, running in llama.cpp and it works GREAT for small python projects, so far.
It was very easy to connect the harness to the local model and it seems to run quite fast, compared to other harnesses that I have tried.
I've been thinking that they should design the DeepSeek harness to work with other providers since a large use case is to off-load work from an expensive model to DeepSeek, instead of makign everyone hack the harness.
I see that it works with many different providers out of the box and that's a great thing. It also makes it easy for me to build a plugin for the role-model router and have it work properly, so you can route between models automatically. Will be out later today.
I'm curious what peolle are finding with first party vs 3red party harnesses for coding.
Do the first party harnesses really have an advantage when paired with the maker's model?
And that's it, that's the last lab releasing models worth coding with that didn't have a first party harness that its models are trained to use.
I personally really like products with plugin systems: a stable, cohesive co with a rich, extensible ecosystem. You can create products that fiyour exact needs, and even if there's no plugin that meets your requirements, you can build it yourself. At least there's vibe coding.
Just like Obsidian, there's also hot loading.
Tracing what it actually did. Who would have thought that's a good idea, instead of trying to obfuscate everything.
Played around with it.
Does what it says on the tin. Great work.
Problem is I use pi.dev and I don't want to learn a new thing the new things keeping coming. Too much new and unprecidented rate.
As someone who's just getting into the self-hosted game on a M2 Pro MBP with muse-glimmer 30b, what's the difference between something like this and Cline?
We can /model, but when can we /harness?
My approach to harnesses is:
Everything is a skill, backed by a CLI tool that both I and the agent can use and debug.
not very impressed.
in the era of AI, telling me that the core design is a plugin system that can be reloaded and extended easily is just not exciting.
it is something you feel excited 20 years ago back in the 2000s, in 2026, the expectation is agentic capabilities and self improving.
Is there any place to see benchmarks for harnesses (not models)? I'd love to see these things compared.
47mb downloaded, 1.5gb after build, wtf?
I consider my own coding agent bloated at just 1mb (yes 1mb) because it uses postgresql package as db tool, and it works wonders.
* edit 1: Upon further scrutiny, 35 dependencies make up for 1.4gb, what they are for? I don't even see postgres in there so I guess that would be another plugin. 1.5gb of basic functionality?
* edit 2: Most of the time I use the terminal but also developed a web ui for my agent [1] and it is only 20mb with postgres, git, web, file tools, etc I definitely want to know why the bloat
We keep reinventing everything. Eclipse style plugins from a long time back
Awesome, let's read what they have done! I open a new tab.
To install the harness, first use npm...
And tab is closed. No thanks.
Spatiotemporal Composability... CORDIS... sounds like a few doctor who fans in deepseek
Is this better than the bytedance cloudwego/eino Go library?
> It uses an architecture where everything is a plugin
Did they discover Unix pipes?
What sorts of questions can we and can't we ask?
Everything is a plugin? I'm reminded of Eclipse!