I just installed DeepSeek Harness for MacOS and it's a beast. The same great harness as before but now is an app you just click on your dock to run
All settings and workspaces are transferred so you don't lose anything, the only thing it lacks is a way to increase/decrease font size with cmd + and cmd - so I'll work on a plugin for that
I can't be happier
Edit: Asked DSH for a plugin to increase/decrease font and it delivered. Love it. Then asked for a plugin to ring a bell on questions and tasks finished, of course it delivered flawlessly. The same plugin architecture, extensible by the same AI
All LLMs work the same to a certain degree, it's a matter of personal preference, token cost and customization. I asked DSH to build a news aggregator for me and it built a really nice app from 30 news sources via rss feeds for less than 50 cents and presto, reading tech news has never been so gratifying
I feel like I live in another reality from others here; I read people saying things ‘while their jaws are hanging open’ (not said here literally but the feel is the same ; I see it in other threads on HN literally here though); why are we so chuffed with stuff that’s now been one shot for maybe already a year, but definitely the last 6 months? Literally everyone who tried LLMs know this and yet it seems a huge surprise to people here that it is so easy now? And that’s opposite of the people, also very much on HN, who say AI is shit at coding and will not replace humans because humans have to fix its bad code.
It has particularly good observability (the ability to see the full content of every prompt and response and tool call) compared to other harnesses, that's the main thing that stood out to me.
Weird. I don't use the official harnesses by OpenAI or Anthropic but with all others i tried there wasn't a single one that hides any of that. What harnesses did you compare to?
I think what he meant was that it gives you good overall observabiity, over your entire context, kind of how langfuse or these platforms do, but it does so locally. A good philosophy i picked up is how they treat the transcript as the backbone of the product.
Do the same as I did, ask it to develop a plugin to increase/decrease font and you will see results in one single task, dsh-zoom as a plugin, restart and done
Most of this thread seems neutral/skeptical with just that post being obvious. On the other hand the volume of jev threads we have been seeing these past 2 weeks..
It works with lots of providers, including local ones.
After you log in, click on your name/picture in the bottom left, Settings, Models. Has a huge list of models supported in the dropdown, plus a "custom model API" tab where you put in the name you want, the base URL, the protocol, and your API key.
Did I miss something? A quick search leads me to this:
>DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI.
>It is built on an everything-is-a-plugin architecture and powered by Cordis, whose design is described in A Programming Paradigm for Spatiotemporal Composability.
Now, this could simply be an attempt at sarcasm but it's getting really hard to tell. There's a handful of people in this thread saying that DeepSeek Harness is closed source.
https://frontierharness.org/ allegedly this is on the Pareto frontier, though I don't know how good of a benchmark this really is. it seems to focus on one-shot type tasks, whereas the real utility of one harness over another seems to make itself known in long running tasks.
also this is a very limited static snapshot with one model as backend, I wish there were more consistently refreshed and diversified harness benchmarks.
i've seen "Pareto frontier" literally 500 thousand times in the last two weeks or so, and maybe once or twice before that. Can someone please explain what happened recently
I would guess that the article came out because the term was getting popular in AI/LLM circles for measuring LLM cost-for-performance metrics. It was probably riding the meme up the wave, rather than fundamentally triggering it.
So for me, literally I was trying to solve a problem with an AI agent and it threw the term at me. It is a real and useful concept, but my hypothesis is that a certain AI release started using and a bunch of people said "Oh this makes me sound smart/trendy" and started using it every chance they got.
We like to think we aren't token predictors, but the amount of sheer regurgitation I see from us humans makes me wonder (also "commoditize your complement", and so many others)
My suspicion is that getting a big binary with full permissions is the goal here. Harnesses will only run their companion models and will demand your Contacts list.
You don't have to use it. I don't use it. I still think it's really cool, and I've been thinking of trying to take ideas from it and integrate it into my own workflow. In particular, the session visualization tooling.
I wish there was a tool I could use to share my documents, activity, etc. directly with the world's major intelligence agencies.
A marketplace, where the different intelligence agencies could evaluate the value of my digital stuff and then offer me something (like a $5 gift card to Olive Garden), would be really nice.
Being Danish I'd probably rate it China > you > US. Realistically though, I'm probably sending everything to everyone except for you.
I rate the US last because the US government shares information with European governments, so it's the most likely to affect me. I wrote this next to my dishwasher though, and if it's anything like those LG tv's it probably identified what I wrote from the keystroke sounds or something.
I live in China, they already have everything, especially when using appl products. It is basically mandatory here since privacy is a completely different concept here (not sure its even a concept here tbh. Although there are some privacy laws)
Why the fuck would I care? The spooks definitely won't. I'm one person on a planet of 8 billion.
I think people don't appreciate the scale of that. A person is like a pixel on a 4k screen, in a wall that's made of many screens (about 5 in case of my country, about 42 in case of the US). Nobody cares what the single pixel does. It's not even visible unless it's so broken it doesn't change the same way surrounding pixels do.
I have a feeling that all this privacy obsession is actually a reaction to the unconscious feeling of irrelevance at this scale. In their minds, people defy the fact they're just sand for the systems of modern civilization - no, they are the very special ones, each fancies themselves a player among billions of NPCs, and to prove their uniqueness and specialness, they want to hide that fact from other NPCs, and get annoyed at the notion the world may indeed discover they are special.
Nothing else makes sense. Not in the west at least, where the only realistic threat from loss of privacy is getting weirder ads (hint: a real power move is to start with ad blocking and removing ads - and ad-infested media sources - from your life).
Now, loss of control over your daily life and your computing devices, that's another story. But most people don't seem to care about those.
Deep inside they're realizing that the whole privacy thing is overblown obsession, and nobody actually cares about their data, and they wish someone did, but clearly no one does, not even $5 gift card for everything there is to know about one.
In the gemini thread, there was someone (rightfully) impressed, how it gdb'ed onto a kernel module and debugged iouring. All the harnisses i used ran atleast an ACL away from private files and juicy capabilities. Wtf is happening in the community?
Not that long ago, the joke was about grandparents installing every available toolbar into IE and getting hacked because "oh, neat! I want that!" without thinking about it.
Yeah, I know that one, but I actually looked into one and found nothing interesting. Mostly a bunch of paranoid people obsessed with "intelligence agencies" constantly looking through their sock drawers, as if spooks had literally nothing better to do with their time.
Teacher of mine spent almost 20 years getting access to his FBI file through FOIA requests. Apparently it was surprising how much of his life they didn't know about.
Children with no life have never had anything worth keeping private. Even the question of having nothing to hide is foreign if you've never done anything at all.
It's like that parable about God giving us free will. If or when the intelligence agencies manage to create a generation with nothing to hide, what will their purpose be? You might as well take field notes on birds or sheep.
The original title was "DeepSeek Harness Desktop app for MacOS and Windows"
This new title says nothing, we already know deepseek harness from before, this is a new product, an installable app worth differentiating from just "DeepSeek Harness"
This page doesn't emphasize the cordis architecture [1], but that's the most exciting thing about this -- not just yet another harness. It has the potential to make this into something like the emacs of harnesses! I think this might be particularly potent for long-running agents.
It's no that fancy when is put in practical engineering terms. I've spend some time with the paper.
Paper describes such system whch has ability to enable/disable capabilities without process restart. Each plugin must follow specific shape: function to activate, function to deactivate it (both working, mutating the same shared context object), describe provided and required services. This, plus some ideas, like "every provider must outlive consumers when destructred" allows plug-n-play ensuring all dependencies are satisfied before the plugin is activated, lazy activation, and plugin deactivation/destruction.
What makes this great is infinite hack-ability of the interface and agents have become pretty good at writing elisp packages, so basically you can have a workflow exactly how you want.
I wish I had time to make a screencast about this, but If you are a Emacs user and not using Emacs for agentic stuff, you are missing out IMO.
In my time using it with GLM 5.2, I've observed it modify itself to disassemble binaries in search of answers - without me needing to ask it to do so. Very powerful stuff. Here's what it looked like:
◇ reasoning summary
│ The 200MB binary grep with regex backtracking is too slow. Let me write the strings out first (using grep -a with simple fixed patterns is fine; regex with .{0,80} on binary is slow). Better: use SBCL to do
│ a fast binary search, or use tr to split null-separated strings then grep fixed strings. Let me use a quick approach: grep -ao with fixed patterns only (no regex), which was fast before.
│
│ To extract context, I can use SBCL to load the file into memory (206MB is fine) and search for occurrences with bounded context. Let me do it in a Lisp worker.
▸ lisp.eval
(lisp.eval :form "(with-open-file (in \"/nix/store/g7b3p8ypxmdycry9wilil8mwl084kplw-antigravity-cli-1.1.19/bin/agy\" :element-type '(unsigned-byte 8))
(let* ((data (make-array (file-length in) :element-type '(unsigned-byte 8)))
(_ (read-sequence data in))
(haystack (sb-ext:octets-to-string data :external-format :latin-1)))
(flet ((find-all (needle)
(let ((out '()) (start 0))
(loop for pos = (search needle haystack :start2 start)
while pos do (push pos out) (setf start (+ pos (length needle))))
… +10 more lines
│ (with-open-file (in "/nix/store/g7b3p8ypxmdycry9wilil8mwl084kplw-antigravity-cli-1.1.19/bin/agy" :element-type '(unsigned-byte 8))
│ (let* ((data (make-array (file-length in) :element-type '(unsigned-byte 8)))
│ (_ (read-sequence data in))
│ (haystack (sb-ext:octets-to-string data :external-format :latin-1)))
│ (flet ((find-all (needle)
│ (let ((out '()) (start 0))
│ (loop for pos = (search needle haystack :start2 start)
│ while pos do (push pos out) (setf start (+ pos (length needle))))
│ … +9 more lines
lisp.eval cpu 0.205s · real 4.018s
│ Output:
│ settings.json positions: (89311959 90892289 90918553 90945997 90985371 91164238
│ 91164426 91171707 91275241 91297838 91299802 91299923
│ 91629274 96365298 96370345 96398990 96405347 96406738
│ 96414968 96419796 96426018 96429189)
│ --- 89311959
│ oughDependency: NestedType: OutputType: Deprecated: AllowAlias: Annotation: SourceFile: rsa-sha2-512rsa-sha2-256hmac-sha1-96ssh-userauthkey exchangedirect-tcpipForwardAgentGatewayPortsTCPKeepAlivedefault no…
│ --- 90892289
│ ks=%v)conversation file not found in source directory: %sfailed to create cache directory for conv cache: %vBalanced speed and reasoning quality for most tasksPrepared switch to default agent (will %s on ex…
│ --- 90918553
│ on-interface type passed to Type.Implementsreflect.Value.Slice: string slice index out of boundsUse /feedback to share your experience with the team.Transition failed: ForkConversation returned empty IDpers…
│ --- 90945997
… +14 more lines
◇ reasoning summary
│ The binary confirms:
│ - modelProvider in settings.json selects the Gemini API backend with GEMINI_API_KEY.
│ - There's a regex ^(?:gemini-api:|blade:|evergreen:|beyond-api:)(?:.*?/models/(.+)|(?:.*/)?([^/]+))$ — model references with provider prefixes.
│ - "No title command configured in settings.json. Please add a "title" block." — settings keys.
│
│ Now I need to find the settings struct JSON tags for the model field. Let me search for json:"model" again with wider context, and also for the string "title" block and other settings-related keys like
│ "theme".
The binaries are from china. Your data goes to china. The self-updating plugin system is vulnerable to malicious llm provider. Anyone using this is crazy!
Is there a safer alternative? It’s not like I can trust OpenAI since they’ve proven that they’ll steal research from mathematicians. X is run by a sociopath. Anthropic is going to IPO so who can say how long they’ll be trustworthy…
Both Z.AI and xAI have already proven you can't trust them, like at all. I guess other harnesses are in "I don't trust it, but it is convenient" category.
It's open source. Unlike Claude Code your beloved friend which is closed-source and uses steganography along with an astonishing array of telemetry and fingerprinting techniques to conduct intelligence analysis on its users.
The new danger is US administration and the current fascit regime. China has yet not topple a single government abroad, has not bombed a single country and has not tapped the phones of its allies.
Now don't down vote before understanding the definition of a fasict:
"A fascist is a person who advocates for or adheres to fascism, a far-right, authoritarian, and ultranationalist political ideology."
I don't see much of a future for these kinds of intricate harnesses, or harnessing in general for that matter. As models are getting better, harnessing will shrink until they are at the level of vanilla Pi or not even that.
agi is artificial general intelligence. sol5.6, astra, fable are already wildly more intelligent than the average person. we already have agi. however people need to keep the moving target so they have something to talk about, otherwise the voracious appetite for novelty will not be met.
what we don’t have yet are the tools to take full advantage of the agi we do have. they’re coming soon.
AGI is when AI models no longer need any training at all, because they have comprehension, meaning human science finally developed artificial comprehension, which we have not yet. Calling what we have now AGI is sycophant talk.
My version of this theory is that we already have AGI, but Altman/Amodei/Zuck keep asking it for an infinite money glitch and it keeps (correctly!) saying "no money, humanity is fucked given the trajectory, and you in particular will be fucked once the general public decides that you're the scapegoat". A/A/Z decides the AGI is wrong of course, so obviously dumping another billion into training or reinforcement or tagging or whatever is the next step.
Exactly, sometimes when reading Hacker News I get the impression this thing is a cult. Luckily there's still some dissident voice around to keep it interesting
drives me up the wall, this shifting definition of AGI; folks, when we no longer need to train AI models at all because we have finally developed artificial comprehension is AGI. This propaganda imposed definition is for fools and sycophants.
> As models are getting better, harnessing will shrink
That can be true only for locally hosted models. The more supplied tools can do, the less data has to be exchanged with OpenAI/Anthropic servers. So bad harness means both higher lag and token usage.
Still in preview, but I think this update is only meant to simplify the installation process.
Not providing DSH as a simple to install package resulted in an unknown third party packaging it with some modifications and SEO the hell out of it to always come on top in Internet searches (even above DS):
deepseekharness[.]io
Could be just an ambitious engineer, but could equally easily be ran by cyber criminals or NSA.
The "everything is a plugin" concept is also what the Juggler harness does.
I'm not yet sure if it is really sensible long-term, but it is definitely now while we're still trying to figure out what exactly we want and need from LLMs and harnesses.
FWIW, making juggler plugin based feels quite normal to me, coming from the world of audio software, where that entire industry is still based around many thousands of VST/AudioUnit plugins which run native un-sandboxed code in the host apps, with access to absolutely anything they want. And it's somehow all just kind of been OK..
DeepSeek harness is the OpenCode (better than OpenCode) of the web/desktop medium.
Good things about it are that it is extremely lightweight and fast. The communication between sub agents is two way in that a sub agent can midway send a message to parent and the parent can send a message midway to change the course of action of a sub agent and while this is happening, you can sitll continue talking to the model on the main thread.
Downside of DeepSeek is that it is constantly in flux which is understandable and they make it very clear themselves that the breaking changes are to be expected.
it's a wrapper around a model that actually executes commands from text input. Claude Code is an harness : the underlying model is Claude (with the version of your choice), Claude produce text output like `grep -in "error" server.log` , Claude Code actually execute the code in your shell and return the output to the model
Anyone know if this works with local models or only the APIs for DeepSeek? I'm assuming it's with the API, but I'm looking for cutting edge local harnesses.
If you're using Mistral Vibe installed on your computer, that is a harness. Vibe is one of many AI harnesses.
In general, a harness is something that lets your AI use tools on your computer and lets it work independently while you walk away. (Very simplified definition, not all harnesses are that automated.)
It's a long time since I tried Vibe, but when I did it was one of the worst. Back then it had broken support for MCP or computer use and other tools people now regard as near essential in AI. Hopefully Vibe is better now. Sadly Mistral are a long way behind.
Mistral Vibe lets you use GLM-5.3 by default now. So it's a very capable model, for a cheap monthly fee. The Web version is excellent : tried yesterday on a simple search, results were way more informative than ChatGPT.
But yes, it's sad that they don't release models anymore.
I've been with Mistral for a while now. Their hosted GLM-5.3 has been a game changer in Vibe CLI over the last couple of weeks. The Vibe CLI harness has been improving for a while, but is still a little sluggish and I feel lacks token minimisation strategies. I recently changed to Maki and have been reasonably satisfied so far.
However, I also use the Mistral Vibe chat in my browser. I think they only switched to GLM-3.5 here a couple of days ago (https://docs.mistral.ai/resources/changelogs). It'd be nice if Mistral had a desktop client, but I have Goose installed which I have no real complaints about.
Captcha had me ruffled at the onset as I am slightly visually impaired and couldn't pick out the somewhat grainy images. It seems to be the standard now.
My problem with plugins is that I don't want to install stuff from third parties. It might be malware, become unsupported, or generally have a lower quality. I rather have a curated and polished feature set out of the box.
It's nice for having your AI create your own plugins. But I haven't really found the need yet.
A plugin system in no way impedes having a set of curated and polished features out of the box. It simply allows the addition of new features to have the same status and be part of the same class as existing features. It's an architectural and design property of the system, not a property pertaining to quality.
In your case you can stick to installing plugins written by first parties or write your own plugin for specific use cases that you come across. You don't need to have any use cases for plug ins right this moment, but perhaps at some point if you find yourself repeating the same manual process over and over again, you can write a plugin to automate a substantial portion of it.
Why does every lab producing an LLM want to have their own harness? Lock-in? Is there any other advantage for the lab?
I'm using Cursor (it's the only way my company allows us to use Grok) and OpenChamber (when using GPT, Muse Spark and others) and I'm happy. If I were to use DeepSeek, I'd use it through OpenChamber too.
deepseek direct is much more cheaper than going through openrouter, openrouter price arbirtages (of course) and has a top up fee and does not pass the discounts that deepseek offer which is 50%+ discount when using it on off hours from beijing time.
meaning USA usage is 12 hours off so you can take advantage off off peak pricing
of course you can ask your ai flavor to tell you this but:
open router 5% deposi fee, none direct
deepseek does prompt caching for token discounts, openrouter not as easily but there
US daytime usage = automatic 50% discount direct w/ deepseek.
Seems like a category error: you can use DSH with openrouter, in fact it ships built-in support using the Pi LLM SDK.
I think DSH is a pretty well-built vanilla harness with a nice web UI (I use it in a pinned browser tab). I much prefer this to dealing with TUI clipboard/scroll jank, and it works just as well over SSH if you simply forward the port. They've genuinely thought about the architecture, and made some effort towards sandboxing the LLM's shell. Yes this is a saturated area, but they've made a nice version of the thing everyone is making, without trying to lock it down to their API. Maybe give it a try?
When my family moved to a house with a dedicated playroom, my wife and I were dreaming of getting the kid’s toys out of the living room. A friend told us “Either they play where you live or you live where they play.” I’ve found that to be true. Do you find a yolo sandbox like this has a similar tradeoff? Like, you don’t have to approve permissions in yolo but you’re constantly copying stuff to and from the sandbox. In the end maybe your stuff ends up “living” in the sandbox.
> The desktop application is an Electron shell around the complete dsh Web application.
I find it hilarious that all the AI harnesses are Electron crap. These are such simple UIs, it seems like it would be trivial to have the AI itself write native frontends.
Way too short sighted. You're still operating under the (mistaken) beliefs that
1. The write terrible unmaintainable code
2. They can't just improve any code retrospectively
3. You're going to be maintaining it
I am now writing home projects where I am actually "vibe coding", to an extent. In my job, I can't do this because other people still review my work and will complain about things that don't match our human invented patterns for software development, and more importantly, I am actually responsible for this code and if it goes wrong our firm goes bust.
But, I genuinely mean it when I say that we are not far away from the point where verifications (functional, non-functional), as long as they're comprehensive enough, will be all that is required, and I don't expect to be reading the nitty gritty details of PRs in a few years time.
It is only a mess if it doesn't work correctly in a functional and non-functional sense. All other coding conventions are invented to make it easier for humans to maintain code. Agents don't care, they will happily chug away pressing the virtual keys endlessly, changing whatever the previous agent left.
AIs do a good job of maintaining their own code, including refactoring, increasing test coverage, ensuring documentation doesn't become stale, updating dependencies. They've been extensively trained to perform pretty much all of the functions of maintaining codebases and they do it far more diligently than most people.
You still can't beat the speed to a functional and usable product. Most users don't mind the 1G download.
I am writing a rust native UI using GPUI but now I need a markdown viewer and editor and also want diagrams. I might end up embedding a webview (tauri style) after all.
Oddly enough, using agent harnesses is what has made it possible for us to ship native macOS, Qt (for Linux), CLI, and Windows clients. The burden of having multiple clients just isn’t that bad now.
Yeah, I get that form follows function, but I’m hoping for more native features of quality control, some sort of dashboard to catch slop. Also looking forward to more domain-specific harnesses.
Did you look into the workflow tools? The agent can write a TypeScript program that encodes a subagent graph, so you can codify a review workflow and ensure it's actually followed. Kind of neat, I think other providers also have these now, but it was new to me, and interesting to see how it worked.
Adding dashboards etc is probably where the "everything is a plugin" starts to pay off. They ship an agent preset for working on the harness itself.
On-topic: I like DeepSeek Harness quite a bit, but the problem with "Everything Is A Plugin" is that when this includes core functionality, you still have to maintain downstream patches for those core plugins if you want to tweak existing behaviour. I currently have ~25 downstream commits and 0 new plugins.
I wish I caught this before my first command, disabled it but not sure what it sent off :c but this is something to enable. It's visible in settings as well, but it does not match what you set on the web portal first even after syncing it.
not surprising. their approaches are generally leaning on gathering as much user data as possible, e.g. the paid API can't opt out training. I'd always double-check before using any service from them.
Yeah, I am actually ok with this. I'm selective about what goes into it, and the training goes into better open-weight models. From what I've seen the DSH desktop telemetry is actually telemetry (like VS Code), not uploading all your files.
Might have been unclear from what I posted originally, but the `session-log-deepseek` is attaching session transcripts to DeepSeek API inference requests, so it's just a full version of what already needs to go in the messages array, with compaction expanded etc.
If you run `dsh web` against a local model then nothing goes to DeepSeek by default (except the web_search tool uses their API, but there has to be a backend somewhere). I think that's commendable. The regression on the desktop app is disappointing.
CCP really doesn't care about policing what I do or think and doesn't have the reach to enforce anything. I can think of a lot of topics that I'd never ask a US-based chatbot that works with the US government, that I'd be fine with the CCP reading instead.
Why does every one of these apps pollute the home directory by default? You'd think AI would be able to write the code to make the app fucking behave properly and store its settings in the platform appropriate way.
I'd like the model provider and harness provider to not be the same company.
This gives more power to the consumer and ensures they can freely switch the components according to their needs. AI market is already full of power imbalance and anti competitive potential, no reason to increase that even more.
[OP] Kuyawa | 18 hours ago
All settings and workspaces are transferred so you don't lose anything, the only thing it lacks is a way to increase/decrease font size with cmd + and cmd - so I'll work on a plugin for that
I can't be happier
Edit: Asked DSH for a plugin to increase/decrease font and it delivered. Love it. Then asked for a plugin to ring a bell on questions and tasks finished, of course it delivered flawlessly. The same plugin architecture, extensible by the same AI
johng | 18 hours ago
I'm not sure what you mean by it's a beast.
[OP] Kuyawa | 17 hours ago
pylotlight | 17 hours ago
zarmin | 17 hours ago
konart | 16 hours ago
zarmin | 2 hours ago
anonzzzies | 15 hours ago
johng | 2 hours ago
cjbprime | 16 hours ago
inexcf | 13 hours ago
adammarples | 10 hours ago
dante54 | 9 hours ago
logicchains | 14 hours ago
ycui7 | 17 hours ago
[OP] Kuyawa | 17 hours ago
sroerick | 15 hours ago
fakwandi_priv | 15 hours ago
sroerick | 15 hours ago
dudisubekti | 15 hours ago
Check GPT-6 thread for example, now that's Astra-turfed lol.
gigatexal | 15 hours ago
hypfer | 15 hours ago
chvid | 13 hours ago
eru | 12 hours ago
rheisen_ | 15 hours ago
(blackbear.app was developed entirely by a custom harness architecture)
proc0 | 14 hours ago
BrentOzar | 6 hours ago
After you log in, click on your name/picture in the bottom left, Settings, Models. Has a huge list of models supported in the dropdown, plus a "custom model API" tab where you put in the name you want, the base URL, the protocol, and your API key.
JKCalhoun | 9 hours ago
IGNORE: Now, if only DeepSeek Harness were open source…
(EDIT: DeepSeek Harness is open source—I missed it).
orangeboats | 9 hours ago
>DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI.
>It is built on an everything-is-a-plugin architecture and powered by Cordis, whose design is described in A Programming Paradigm for Spatiotemporal Composability.
https://github.com/deepseek-ai/deepseek-harness
Now, this could simply be an attempt at sarcasm but it's getting really hard to tell. There's a handful of people in this thread saying that DeepSeek Harness is closed source.
scuppernong | 17 hours ago
also this is a very limited static snapshot with one model as backend, I wish there were more consistently refreshed and diversified harness benchmarks.
polotics | 15 hours ago
hypfer | 15 hours ago
It happens more often than you think, especially when VRAM is limited!
dgellow | 10 hours ago
cdelsolar | 14 hours ago
vrganj | 14 hours ago
zmmmmm | 12 hours ago
scuppernong | 2 hours ago
233mhz | 14 hours ago
JKCalhoun | 9 hours ago
ordersofmag | 9 hours ago
zinoc | 13 hours ago
pu_pe | 12 hours ago
thethirdone | 5 hours ago
[0]: https://trends.google.com/trends/explore?date=all&q=pareto%2...
[1]: https://books.google.com/ngrams/graph?content=pareto+frontie...
dualvariable | 2 hours ago
zug_zug | 10 hours ago
We like to think we aren't token predictors, but the amount of sheer regurgitation I see from us humans makes me wonder (also "commoditize your complement", and so many others)
JKCalhoun | 9 hours ago
zihotki | 13 hours ago
jamienk | 17 hours ago
neya | 17 hours ago
webXL | 17 hours ago
I'd probably check it out if it was actually called 'YAH Harness' and not [insert Chinese AI lab here] Harness
Zambyte | 16 hours ago
djriley | 15 hours ago
quyleanh | 17 hours ago
With the preview of official app, hope team will add remote connection soon.
Scarbutt | 17 hours ago
https://github.com/deepseek-ai/deepseek-harness/blob/master/...
https://github.com/deepseek-ai/deepseek-harness/blob/master/...
throwa356262 | 15 hours ago
It currently does not work with Bun and it is somewhat complicated to run it in a remote docker container due to the its weird security model.
ElProlactin | 17 hours ago
A marketplace, where the different intelligence agencies could evaluate the value of my digital stuff and then offer me something (like a $5 gift card to Olive Garden), would be really nice.
kvirani | 17 hours ago
rapind | 16 hours ago
npn | 16 hours ago
Sabinus | 14 hours ago
Quothling | 14 hours ago
I rate the US last because the US government shares information with European governments, so it's the most likely to affect me. I wrote this next to my dishwasher though, and if it's anything like those LG tv's it probably identified what I wrote from the keystroke sounds or something.
thenthenthen | 14 hours ago
TeMPOraL | 14 hours ago
Why the fuck would I care? The spooks definitely won't. I'm one person on a planet of 8 billion.
I think people don't appreciate the scale of that. A person is like a pixel on a 4k screen, in a wall that's made of many screens (about 5 in case of my country, about 42 in case of the US). Nobody cares what the single pixel does. It's not even visible unless it's so broken it doesn't change the same way surrounding pixels do.
I have a feeling that all this privacy obsession is actually a reaction to the unconscious feeling of irrelevance at this scale. In their minds, people defy the fact they're just sand for the systems of modern civilization - no, they are the very special ones, each fancies themselves a player among billions of NPCs, and to prove their uniqueness and specialness, they want to hide that fact from other NPCs, and get annoyed at the notion the world may indeed discover they are special.
Nothing else makes sense. Not in the west at least, where the only realistic threat from loss of privacy is getting weirder ads (hint: a real power move is to start with ad blocking and removing ads - and ad-infested media sources - from your life).
Now, loss of control over your daily life and your computing devices, that's another story. But most people don't seem to care about those.
ElProlactin | 13 hours ago
Is that really too much to ask for?
TeMPOraL | 14 hours ago
postsantum | 17 hours ago
Reminds me of some magic device from childrens book, don't remember its name, was like a glass ball where you could see what other people are up to
throwawayqqq11 | 15 hours ago
itgoon | 14 hours ago
rootsudo | 5 hours ago
TeMPOraL | 14 hours ago
tobi_bsf | 16 hours ago
Shin-- | 16 hours ago
smitty1110 | 16 hours ago
cheonic4040 | 16 hours ago
So will western countries.
At least chinese offers 95% of the quality at 10% of the price.
kindablissy | 15 hours ago
Bluestein | 16 hours ago
kkarpkkarp | 14 hours ago
oh, good old times of PRISM, rightly bygone. Of course they don't do this now.
noduerme | 14 hours ago
noduerme | 14 hours ago
Children with no life have never had anything worth keeping private. Even the question of having nothing to hide is foreign if you've never done anything at all.
It's like that parable about God giving us free will. If or when the intelligence agencies manage to create a generation with nothing to hide, what will their purpose be? You might as well take field notes on birds or sheep.
[OP] Kuyawa | 17 hours ago
This new title says nothing, we already know deepseek harness from before, this is a new product, an installable app worth differentiating from just "DeepSeek Harness"
nezhar | 15 hours ago
ssivark | 17 hours ago
[1] https://arxiv.org/abs/2608.25512
podgorniy | 13 hours ago
Paper describes such system whch has ability to enable/disable capabilities without process restart. Each plugin must follow specific shape: function to activate, function to deactivate it (both working, mutating the same shared context object), describe provided and required services. This, plus some ideas, like "every provider must outlive consumers when destructred" allows plug-n-play ensuring all dependencies are satisfied before the plugin is activated, lazy activation, and plugin deactivation/destruction.
dgellow | 10 hours ago
ta20240528 | 10 hours ago
nicce | 10 hours ago
dgellow | 9 hours ago
4b11b4 | 9 hours ago
gnufied | 8 hours ago
What makes this great is infinite hack-ability of the interface and agents have become pretty good at writing elisp packages, so basically you can have a workflow exactly how you want.
I wish I had time to make a screencast about this, but If you are a Emacs user and not using Emacs for agentic stuff, you are missing out IMO.
evanjrowley | 6 hours ago
Would you be interested in a self-modifying harness written in Common Lisp? If so, then please check out Autolith: https://github.com/lambda-symbolics/autolith/
In my time using it with GLM 5.2, I've observed it modify itself to disassemble binaries in search of answers - without me needing to ask it to do so. Very powerful stuff. Here's what it looked like:
kgwgk | 4 hours ago
(The name sounds a bit like autolytic, evoking self-harm. Which may be appropriate after all.)
ssivark | 3 hours ago
accountrequired | 16 hours ago
sampullman | 16 hours ago
glenpierce | 15 hours ago
Good4boothee | 12 hours ago
And Pi is considered the neutral choice: https://news.ycombinator.com/item?id=49926069
browningstreet | 8 hours ago
jluysvi | 15 hours ago
qurren | 15 hours ago
What are they going to do? They have no jurisdiction over me.
harvey9 | 15 hours ago
wickedsight | 14 hours ago
nextaccountic | 15 hours ago
Sha1rholder | 15 hours ago
bedane | 15 hours ago
yjh0502 | 15 hours ago
bedane | 6 hours ago
latentsea | 15 hours ago
wg0 | 14 hours ago
Now don't down vote before understanding the definition of a fasict:
"A fascist is a person who advocates for or adheres to fascism, a far-right, authoritarian, and ultranationalist political ideology."
China is none of that.
edelhans | 12 hours ago
StrauXX | 15 hours ago
throwa356262 | 15 hours ago
My wild theory is that we already have AGI level models, but we are not yet using them correctly.
StrauXX | 15 hours ago
throwa356262 | 15 hours ago
010ED67913 | 14 hours ago
what we don’t have yet are the tools to take full advantage of the agi we do have. they’re coming soon.
bsenftner | 10 hours ago
ozozozd | 14 hours ago
/s
Octoth0rpe | 15 hours ago
grim_io | 14 hours ago
ozozozd | 14 hours ago
Kind of like how we all fail to see the bearded guy in the sky.
puelocesar | 13 hours ago
bsenftner | 10 hours ago
surgical_fire | 13 hours ago
I think that harnesses for LLMs are as important as IDEs for languages.
Good4boothee | 12 hours ago
That can be true only for locally hosted models. The more supplied tools can do, the less data has to be exchanged with OpenAI/Anthropic servers. So bad harness means both higher lag and token usage.
kindablissy | 15 hours ago
throwa356262 | 15 hours ago
Not providing DSH as a simple to install package resulted in an unknown third party packaging it with some modifications and SEO the hell out of it to always come on top in Internet searches (even above DS):
deepseekharness[.]io
Could be just an ambitious engineer, but could equally easily be ran by cyber criminals or NSA.
blain | 15 hours ago
hypfer | 15 hours ago
I'm not yet sure if it is really sensible long-term, but it is definitely now while we're still trying to figure out what exactly we want and need from LLMs and harnesses.
bsenftner | 10 hours ago
julesrms | 5 hours ago
wg0 | 14 hours ago
Good things about it are that it is extremely lightweight and fast. The communication between sub agents is two way in that a sub agent can midway send a message to parent and the parent can send a message midway to change the course of action of a sub agent and while this is happening, you can sitll continue talking to the model on the main thread.
Downside of DeepSeek is that it is constantly in flux which is understandable and they make it very clear themselves that the breaking changes are to be expected.
jdw64 | 14 hours ago
contravariant | 14 hours ago
metchio | 14 hours ago
vrganj | 14 hours ago
cpursley | 14 hours ago
proc0 | 14 hours ago
barcoder | 14 hours ago
JamesMcMinn | 13 hours ago
lazyjones | 10 hours ago
maelito | 14 hours ago
233mhz | 14 hours ago
SyneRyder | 13 hours ago
In general, a harness is something that lets your AI use tools on your computer and lets it work independently while you walk away. (Very simplified definition, not all harnesses are that automated.)
It's a long time since I tried Vibe, but when I did it was one of the worst. Back then it had broken support for MCP or computer use and other tools people now regard as near essential in AI. Hopefully Vibe is better now. Sadly Mistral are a long way behind.
maelito | 13 hours ago
But yes, it's sad that they don't release models anymore.
Thanks for the definition !
phillc73 | 10 hours ago
However, I also use the Mistral Vibe chat in my browser. I think they only switched to GLM-3.5 here a couple of days ago (https://docs.mistral.ai/resources/changelogs). It'd be nice if Mistral had a desktop client, but I have Goose installed which I have no real complaints about.
browningstreet | 8 hours ago
Xunxi | 14 hours ago
Captcha had me ruffled at the onset as I am slightly visually impaired and couldn't pick out the somewhat grainy images. It seems to be the standard now.
wren6991 | 5 hours ago
dude250711 | 13 hours ago
d2kx | 12 hours ago
flohofwoe | 12 hours ago
The webpage is a completely unusable stuttery mess on my Android phone (at most 5fps).
Jeeetendra | 11 hours ago
Lucasoato | 9 hours ago
browningstreet | 8 hours ago
redox99 | 9 hours ago
It's nice for having your AI create your own plugins. But I haven't really found the need yet.
JKCalhoun | 9 hours ago
233mhz | 7 hours ago
Then ask the model of your choice to implement the plugin for you, I created 3 this morning just for myself
Maxatar | 4 hours ago
In your case you can stick to installing plugins written by first parties or write your own plugin for specific use cases that you come across. You don't need to have any use cases for plug ins right this moment, but perhaps at some point if you find yourself repeating the same manual process over and over again, you can write a plugin to automate a substantial portion of it.
gizmodo59 | 9 hours ago
juujian | 9 hours ago
zxspectrum1982 | 9 hours ago
I'm using Cursor (it's the only way my company allows us to use Grok) and OpenChamber (when using GPT, Muse Spark and others) and I'm happy. If I were to use DeepSeek, I'd use it through OpenChamber too.
* OpenChamber is a GUI for OpenCode.
conception | 9 hours ago
api | 9 hours ago
I'm reluctant to run harnesses from US companies, and prefer to use them in a VM. No damn way I'm installing one from China.
For coding I like to use Zed right now. Pretty solid integration of an AI harness and a good conventional code editor. Might give Pi another look.
rootsudo | 8 hours ago
rootsudo | 8 hours ago
deepseek direct is much more cheaper than going through openrouter, openrouter price arbirtages (of course) and has a top up fee and does not pass the discounts that deepseek offer which is 50%+ discount when using it on off hours from beijing time.
meaning USA usage is 12 hours off so you can take advantage off off peak pricing
of course you can ask your ai flavor to tell you this but: open router 5% deposi fee, none direct deepseek does prompt caching for token discounts, openrouter not as easily but there
US daytime usage = automatic 50% discount direct w/ deepseek.
vanchor3 | 4 hours ago
wren6991 | 5 hours ago
I think DSH is a pretty well-built vanilla harness with a nice web UI (I use it in a pinned browser tab). I much prefer this to dealing with TUI clipboard/scroll jank, and it works just as well over SSH if you simply forward the port. They've genuinely thought about the architecture, and made some effort towards sandboxing the LLM's shell. Yes this is a saturated area, but they've made a nice version of the thing everyone is making, without trying to lock it down to their API. Maybe give it a try?
franze | 8 hours ago
https://aifcc.franzai.com/
delijati | 8 hours ago
omalled | 8 hours ago
trollbridge | 4 hours ago
rayiner | 8 hours ago
I find it hilarious that all the AI harnesses are Electron crap. These are such simple UIs, it seems like it would be trivial to have the AI itself write native frontends.
benterix | 8 hours ago
gavmor | 7 hours ago
munksbeer | 5 hours ago
But, I genuinely mean it when I say that we are not far away from the point where verifications (functional, non-functional), as long as they're comprehensive enough, will be all that is required, and I don't expect to be reading the nitty gritty details of PRs in a few years time.
It is only a mess if it doesn't work correctly in a functional and non-functional sense. All other coding conventions are invented to make it easier for humans to maintain code. Agents don't care, they will happily chug away pressing the virtual keys endlessly, changing whatever the previous agent left.
Maxatar | 4 hours ago
cromka | 7 hours ago
rayiner | 7 hours ago
vini | 7 hours ago
heliosAtwork | 4 hours ago
I am writing a rust native UI using GPUI but now I need a markdown viewer and editor and also want diagrams. I might end up embedding a webview (tauri style) after all.
trollbridge | 4 hours ago
ravila4 | 7 hours ago
wren6991 | 7 hours ago
ravila4 | 3 hours ago
wren6991 | 3 hours ago
Adding dashboards etc is probably where the "everything is a plugin" starts to pay off. They ship an agent preset for working on the harness itself.
wren6991 | 7 hours ago
If like me you're mildly bothered by this, add these lines to $DSH_HOME/cordis.patch.yml before first startup ($DSH_HOME is ~/.dsh by default):
On-topic: I like DeepSeek Harness quite a bit, but the problem with "Everything Is A Plugin" is that when this includes core functionality, you still have to maintain downstream patches for those core plugins if you want to tweak existing behaviour. I currently have ~25 downstream commits and 0 new plugins.rootsudo | 6 hours ago
goobatrooba | 6 hours ago
wren6991 | 6 hours ago
jimmydoe | 6 hours ago
declan_roberts | 6 hours ago
wren6991 | 6 hours ago
Might have been unclear from what I posted originally, but the `session-log-deepseek` is attaching session transcripts to DeepSeek API inference requests, so it's just a full version of what already needs to go in the messages array, with compaction expanded etc.
If you run `dsh web` against a local model then nothing goes to DeepSeek by default (except the web_search tool uses their API, but there has to be a backend somewhere). I think that's commendable. The regression on the desktop app is disappointing.
miroljub | 5 hours ago
Why do you suppose everyone would rather share data with Mosad and CIA than with CCP?
In China, my data is safe. Safe from you, safe from the CIA, safe from Mosad.
cheonic4040 | 4 hours ago
Anthropic also shares your personal data with the US government.
At least DeepSeek provides 95% of the performance for 10% of the cost.
dualvariable | 3 hours ago
wpm | 5 hours ago
wren6991 | 5 hours ago
trollbridge | 4 hours ago
m3kw9 | 7 hours ago
NooneAtAll3 | 6 hours ago
website doesn't work without webgl :/
thih9 | 5 hours ago
This gives more power to the consumer and ensures they can freely switch the components according to their needs. AI market is already full of power imbalance and anti competitive potential, no reason to increase that even more.
wren6991 | 4 hours ago