Interesting to see they shipped an "anti-slop" taste skill:
> description: Anti-slop frontend skill for landing pages, portfolios, and redesigns. The agent reads the brief, infers the right design direction, and ships interfaces that do not look templated. Real design systems when applicable, audit-first on redesigns, strict pre-flight check.
> - *PREMIUM-CONSUMER PALETTE BAN (mandatory, second-most-recurring AI-tell):*
- For premium-consumer briefs (cookware, wellness, artisan, luxury, heritage craft, DTC home goods, etc.)...
- Backgrounds: `#f5f1ea`, `#f7f5f1`...
> Landing pages and portfolios are *visual products*. Text-only pages with fake-screenshot divs are slop.
> Em-dash (—) is COMPLETELY banned. It is the LLM's signature stylistic crutch and it is the #1 visual Tell in production tests. There is no "limited use" allowance, no "natural language frequency" allowance, no "in body copy is fine" allowance. None.
Probably not, because now you'll have all of the people who are worried about this ensuring their AI generated text never ever uses em-dashes, and then that will be the new tell.
Or we can just be more concerned with whether something is well written and well presented and accept that sometimes that is going to be AI written text.
> Or we can just be more concerned with whether something is well written and well presented and accept that sometimes that is going to be AI written text.
Slop is slop, it doesn’t matter if it was human or LLM written. If a code comment or Slack message is 200 words long but still makes no fucking sense then it’s slop. A message can be 10 words long and riddled with typos but if the message is understandable it’s not slop.
I've been posting my writing on the internet since 1990, from a Mac where we all learned how to make en and em dashes, umlauts, accents, degree symbols, proper single and double quotes, and so forth from Day #1 and I have absolutely no plans to stop using em dashes.
Though I agree with GP: models and harnesses are being updated so hard to avoid any use of em-dashes that soon it will be a tell of human writing. The the pendulum will swing back and forth, forever.
I have been seeing people use it more who I am fairly confident are not otherwise using AI to write public-facing communications. It's almost like there's a minor movement to try and reclaim it. Either that or it's somehow getting past my AIdar.
This quote is so egregiously stupid, I'm instantly disinterested in this product now. No reason to assume the rest isn't the same applied cretinism as the quoted sentence.
It's notable that this appears to be used to generate frontends that look more authentic than the typical AI generated site, not as a test for what's acceptable from a human contributor.
I suppose all the exotic unicode characters are dead now. Anything that's too annoying for a human to type but trivial for AI to generate is probably done for. I used to enjoy using those characters. Sigh.
All due respect to the YC folks, but having a skill with 22,069 tokens is a major skill issue. Yes, ironic.
My slop control skill is a thousand tokens. Biggest problem I see with the skill is that everything is prompted via negativa. smh. sorry to be judgmental but it's hard to trust a harness that comes shipped with a skill like this.
The YC folks know engineering hype, not engineering. Most of them have been out of the game so long and focused on other things they've forgotten their fundamentals. You can hear this explicitly in shit Gary says in a lot of videos, it's directionally correct but detached from fundamentals.
Doesn’t this just lead to a new “basin of tastelessness” that, sure, looks different from current slop, but is itself just eventually slop all the same?
Can someone give me an example where this is truly and uniquely useful? I've seen so many of these things and I can't tell if part of it is like mostly for enabling nontechnical folks to do more things or if it's some unlock and additive value. At end of day beneath it all things are just prompts, and then you can provide tools and context, but even then that's not even something I've found that useful to keep because context windows are still limited and bias propagations often need to be constantly corrected, especially when digital artifacts are not perfect representations of the real world.
A video would've been nice to walk through how it's actually used. I don't mean just a walkthrough after a fresh install, but an instance that's actively used.
The hardest problem in multiplayer agents, at least for us, has not been the agent loop. It is scoping and QM's per-person scopes plus shared rooms is a sane answer for a company-wide assistant.
I build in the adjacent lane, AQ (aq.dev), a multiplayer coding harness where teams run Claude Code and Codex together), so seeing YC ship "a multiplayer agent harness for work" is validating and a little surreal.
Still figuring it out, but it's been really convenient to have an always-on agent that has access to internal systems and can be triggered by webhooks. Some examples of what we use it for:
- automatically fixing simple CI failures
- getting production alerts and automatically creating RCAs and a fix PR
- periodically checking slow DB queries and finding ways to speed them up.
- creating charts to answer one-off questions about our data
I've tried using it as an on-the-go coding agent as well, but found I prefer more interactive agents, so I can see what the code looks like.
I think the interesting challenge isn’t running agents, it’s reviewing their work. The more code agents produce, the more important provenance, review ergonomics, and trust become. I also suspect repository platforms will need to evolve there over the next few years.
At least for me, I have a couple dozen years of DB experience but robot, given performance metrics, can get really close to optimal on a tactical level (single query or pattern of queries) but can't yet do the full normalize/denormalize level of improvements without supervision. But really solid if you have one misbehaving query and give it explain analyze access on a read only account
Hermes is huge and packed with features you probably don't need. I prefer smaller one I can extend as necessary, there are so many on github now and it is fun to test them but have been impressed with dirge (https://github.com/dirge-code/dirge) not affiliated.
I have one reading my second tier RSS feeds and newsletters and giving me news/market updates filtered for things important to me
I'm not contradicting this but offering a contrast, I like Hermes because it simultaneously lowers barrier of entry and shows you what possibilities are unlocked by agents. I don't think I would have the time, interest or creativity to jump into the deep end by either extending an existing harness or rolling my own from the start. This also isn't an argument for doing just that, I might do so in the future, but critically only after Hermes has shown me what's possible and my preferences are developed.
I completely agree. I started with aichat[0] before the current agent trend and hit all kinds of bumps implementing agentic loops on my own. Then goose[1] showed me what a whole team working toward the same idea could do right before the official Claude Code harness which had all the bells and whistles. Now I know better what I want and its mostly less ram usage and a small set of primitives.
I still use Claude code (and codex and other big contenders) because they know what they are doing and innovate in ways I don't want to miss. And sometimes they are better at tasks.
I feel the same but my preference is for Cecli (cecli.dev)
It does what I need it to do and since I invested so much time in setting it up and even contributing to development it is my go to for coding and even managing my VPS as well as business tasks
I had the same line of thought and spent a while with nanoclaw before realizing that adding the features I want back in would have made future updates too painful. I ended up switching to Hermes and aggressively disabling tools/skills and it's been pretty fine so far. I got more use cases set up than I did in my time with nanoclaw.
Totally fair, I still use Claude code for tasks since it is so polished. I think I just dont have that complicated of use cases so in the end performance should reflect my lightweight needs as the priority.
Hermes is what I was using but I still found it annoying I often wanted to operate 1-2 levels deeper.
Yesterday I just decided to try writing my own version (100% just for me, not open source, no monetization plan, incredibly custom) and I've been enjoying working on it so far (I know, I know, it's been a day, honeymoon period and all that).
Part of it is I like building software (even if I'm not writing every line) and part is I like having full control. Turns out (and many people have said this) the basic agent loop really isn't all that special. There are a million levers to pull and what not outside the base loop but that the fun part for me. Trying out different ways to add on to the core concept.
I'm really enjoying being untethered for things like "how will I monetize?" or "how do I make this generic so others can use it?". If I need functionality I just add it in, I don't need to make it infinitely pluggable, etc.
All that said, I'm thankful to things like nanoclaw and then Hermes for exposing me to the core ideas. I just want to put my own spin on it.
I didn't have good experience with Hermes. Vellum.ai was better, but unfortunately it had a bug where opencode go providers failed to work for a time, so I instead starting writing my own.
I'm currently using it to help me with my oncall, first responder to our any production alerts. It's not as efficient as coding agent by default, but it's been tremendously helpful to me.
hermes is great to get started with, but it's packed to the gills with stuff you'll probably use one time just to test it. and this eats in to your context so if you're hoping to run it on a lighter-weight local model you'll run in to some trouble. if you go into it planning to customize/thin it out it's solid
So that explains why there is so much low effort cold outreach on LinkedIn from YC founders these days.
It's getting ridiculous the amount of unsupervised agents doing active inbox management on things that should be personal relationship work. I'm so tired of it.
It's fascinating to see new UI primitives and concepts get invented in the LLM era. The sea of creativity makes it hard to even understand most of what each new app does, and nobody describes them well. When I went to the Hermes agent web page, I was left with zero clue about what it did or what it could do. It took a bit of digging to find the right part of the qm page that helped me grok what was going on.
I've become attached to Orca (yc-backed) for managing coding sessions in the past week, but some sort of postgres session db is what's really lacking. So, maybe it's time to try qm.
It is fascinating and I love to see it. Ultimately though, why not build your own? I think that in part is what we're beginning to see-- highly customized and personalized software. I take most of these as inspiration these days and just build my own. Nothing you can't hammer out with a few good Claude sessions.
But why build your own from scratch when you can do it on top one extensible UI platform that already has all of your organizational context, not to mention coding agents already built-in that you can just tell it to build & deploy your personalized workflow softwaree from scratch? ;-)
Because it's fun! I don't "really" use LLM for anything serious yet, but after discovering playing with a custom harness I licked blood. Like going back to edit any message to start a new conversation branch off that, or being able to go back and remove things that turned out to to be dead ends from the history. Not really "useful" [0] for now, I always wondered about that and I was dumbfounded how trivial and tiny a first MVP was. Of course, that's also because it is trivial, it's just chat, it affords none of the usefulness you'd want for any "work", but it's still very interesting.
And for some simple recurring tasks or local housekeeping stuff, having something where I 100% know what files it reads or writes will surely be useful. E.g. if I wanted to watch out for certain topics on HN, I'd rather make something myself that grabs the feed and turns that into a list of titles and topic ID, which is probably 1% of what the front page HTML would be, and then only have the LLM process that output -- rather than telling an LLM to do all that every time. Even if tokens may not be that precious, and the difference in "cognitive performance" not worth speaking of, that would feel way neater to me.
That's an option right? When I say roll your own-- it could be extending another platform. There's definitely different levels. Customization is in reach regardless of where you want to start and stop your stack.
These new experiences are built atop layers and layers of bedrock libraries that we, generally, don't reinvent. Or we reinvent one or two.
What's notable to me is that ux doesn't so far generally have this behavior. That as per this post people just build a new app, a new experience.
I want to believe over time we'll have better composable & malleable ux experiences atop broader platforms for us. That over time the "go it alone" path has other worth ways to innovate that use a more substantial shared base. It's dangerous to go it alone, and doing so equipped with just our wooden sword and some courage and perhaps an LLM wisp is an amazing adventure, but I think the survival rate & impact would be much better if we had more general ux systems that supported better innovation atop them, and if less people did the pure "why not build your own" path.
I understand you take. And I definitely agree that we could use a lot better ux tools than we have right now. Maybe one of these prototypes will lead to that though. I'll look at it more as creative engineering. Not everything is meant to be distributed or consumed. But that doesn't mean that it's not worthwhile to try and prototype different flows and options. I have seen hundreds of harnesses, I've only used a couple-- but they've inspired me to build something that fits me better and works for me. If somebody comes out with something that helps unify the UX experience and provides better building blocks for me to work with, I'll definitely switch to that and rebuild my experience on top of that. But for now this is what we have. Admittedly this reflects my level of experience and my expertise too. I don't have a lot of experience working with building UX frameworks, so I adapt what I can.
People are using the UX systems though, just that it's the agent that is gluing them together now. Under the hood, it's still React plus component libraries like shadcn, lucide, tailwind, axios and so on.
The agents are so good at this, there's literally no point in not doing it that way.
These are all surface veneer but so so so much less than an app. The entire rest of the owl has to be redrawn, and all the guts inside the owl redesigned.
UX does have this, and it has for many years. You might remember Bootstrap way back in the day, and everything looking like bootstrap. These days it seems like everything is based on tailwind under the hood.
Right now shadcn/ui might be the leading candidate but there are plenty including old school bootstrap, material / MUI, a bunch of various flat-themed ui, etc. They often come in the form of a react component library or whatever but they're generally really solid. I also found one in svelte that I can't immediately pull up but it was nice looking too.
It's really hard to focus on the unique aspect of your tool because it runs the risk that people will think it's unfamiliar and therefore not useful to them. That's why a lot of the tool marketing pages look the same even though the tools themselves might be different or innovative
I think we need a new area of study around UI/Agent connection. It can kinda be done with tools, but I'd we need much deeper primitives to allow the UI to inform the Agent and vice-versa. Right now we've just given up, replacing the UI with an Agent message view, but I think that's just because nobody is thinking about how the two can compliment each other
I've been playing around with using a hidden markov model informed by a UI state event stream, with the end state fed into the Agent as a hint on each message turn. Then the Agent can make a tool call to add events to the HMM. This has been really interesting, but I haven't struck the right balance to make it actually feel good for the user yet
Isn't this the direction Claude on the web is going towards? It's a mix of message view on the left and specific interactive UI on the right panel. There is also MCP Apps spec now
Aren't there a ton of products already doing this? Why not just use claude Cowork? Surely they're simpler/better/more featureful/developed than the alternatives here? What advantage does this have? Would love to see a 'QM vs Cowork' comparison!
Because people want to be able to do things like use their own clients of pi or opencode with LLMs they run themselves, such as the just released deepseek v4 flash 0731, not permanently tied to an Anthropic ecosystem of non-open-weight LLMs and pay forever per token.
It feels like the big thing they are touting here is the shared company brain. Not clear to me though, how that brain is developed when each person has their own harness. (I did only skim the docs.)
And it still has the issue of: if the agent is acting as me, then security wise it can do anything I can do. Maybe that's why they are recommending for startups.
People are building state-of-the-art harnesses in private right now that outperform the average developers' harnesses, people using Claude Code, OpenCode, etc.
I love the concept to use different harness frameworks in qm but a true multiplayer harness needs to support other agents and any MCP clients, including Cowork.
Making agents multiplayer is mostly a context problem. You could be using ChatGPT or a Slack bot, or a web interface and the agent needs to know you, your conversations in Cowork etc. so it can enable multi channel collaboration with your agents and your colleagues. We're working on it at https://lobu.ai
Multiplayer agent harness is useful only when the context and AI output does not overwhelm the team space and works asynchronously and background.
If you think about it, that could be just boring job scheduler, isn't it?
A team that pioneered writing small checks for outsized returns. AI is making that venture model a lot less attractive, so now they're building productivity software for the startups they once only invested in.
At least for me, I often have a hard time believing people’s stated reasons for whatever they’re doing.
The substance <> narrative relationship is backwards a lot of the time. Someone does something for a nebulous multitude of reasons, and then post-hoc fits their decision-making into a logical explanation that sounds nice.
There is some psychology research supporting this as well.
I suppose if you see the narrative itself as part of the release, that might be interesting. But most of the time I’d rather just hear plainly and straightforwardly what the thing is.
Or at the very least, I am very accepting of releases which do not include rationalization / narrativization and don’t think it’s required to include.
qm has been a qemu (virtualization) command line tool probably for 20 years or so. The ignorance of these people inside the same field is egregious. (Oh, yes, there can be tools that are named the same in the same field, I get it)
This is the exact same as the breathless faddishness of celebrity news. "Use these seventeen different skincare products and you'll look exactly like Gwyneth Paltrow".
> We take contributions as human-written text, not code — see CONTRIBUTING.md. Describe the change you'd like informally in a .txt or .md file in adrs/, and if we're aligned we'll handle the implementation.
Interesting approach to open source contributions. Closer to feature requests at that point?
I'll need to explore how they're doing org wide context and security for sure. This seems extremely complementary to my own coding tool which currently gives the best AI interface for individuals.
Would be cool if in a few epic tickets i'll have both their org wide architecture AND a productive individual coding interface :)
If it is so not novel, why did I never see it in Enterprise?
Why is their own attempt a barebone mvp?
Besides Dassault Systemes or Palantir with their Ontology product, who was doing anything adjacent?
It's giving typical HN vibes ' you don't need dropbox, just sync to ftp' ;D
looks like an internal tool that yc rushed out the door to minimize any most lost ground to Buzz. That said, I'd be curious what folks think comparing these two tools.
it says
"Each deployment runs in the operator's own cloud account"
But feels like its written to run on one mac/vm and carries same drawbacks of other similar platforms. I'd rather use Hermes/Openclaw for oss or closed managed agents like Tasklet or Prajvis
> Learn your writing voice from past sends, then triage your inbox on a schedule — labels and reply drafts included
Oh grand. The world will be full of agents talking to agents and nothing getting done. So pretty much the same only I can just leave my machine awake to look busy.
i find something a bit funny in an ai project, written by ai, requiring human-written text with specific guidance to not use ai.
"Given that coding agents write most underlying code now, we'd prefer PRs in the form of human-written text. [...] Please do not have AI artificially expand what you'd like to do into a formal proposal."
i think it makes perfect sense and I'll probably adopt the same posture for my open source project.
The idea is that the core dev team is the one with the AI harness. If someone contributes an idea they can just feed it to their AI to implement it and it probably costs 10 minutes of human time, because they trust that their harness will be able to implement the feature competently.
So, because the implementation effort is so low, the only aspect that matters if the quality of the idea. It's way easier to screen human written text than a bunch of code in a PR. If you just give them code, they would not know if you made it with a competent harness. If you gave them a AI written design they would have a lot more to read through to decide whether it's slop. If you just give them an idea it's a lot easier to determine whether it's high quality.
Well, if they're consistent, the ADR directory contains only human written text that drives the underlying development... Aaand it's empty. Well.
I personally would probably have the readme generated based on that directory as the primary document, probably with another `readme_generation_rules.md` in the ADRs directory, and I would be pretty ruthless about disallowing all the slop-adjacent wording.
> I would be pretty ruthless about disallowing all the slop-adjacent wording.
I’d think that trying to play whack-a-mole with slop like that would result in a lengthy prompt, as well as being brittle to future changes.
If you want the prompt to be the source of truth, you’re probably going to have to accept prose that feels like AI, at least until you can regenerate with some future smarter model.
Not even close :) It's really, really good at making natural-sounding prose if you demand it, set clear guidelines (ideally a script it can run with grading, but...), Fable in particular is eerie about mimicking prose style. Try RFC style if you want really clear architecture documents with MUST/SHOULD etc.
It will, of course, be slightly worse at the non-style task (i.e. make the readme worse), but if you care about the style, that's probably an acceptable tradeoff.
Coding agents are extremely useful but often extremely dumb with design. If you do not design the software yourself, you will probably get slop. This was always the core issue with vibe coding.
AI Averages, and is inclined to do average designs and implementations, which in some cases might be an improvement, but long term it creates more to deal with.
Because competent people are evaluating it. In places where the standards are rock-bottom, the AI is leagues ahead of anything they'd produce normally.
i read the emphasized "human-written" part, in combination with that last line, as a blanket restriction (for contributing). but perhaps i read it wrong.
yes, specifically they want purely human-written ones.
i understand why they want their own agent to do the code, and i can see the reasoning behind it. but not allowing ai to format/tidy up/flesh out the proposal is the part that i thought was a little funny.
That's not surprising. I have much the same thing in my latest open-source work[0]. I have a workflow that's working nicely with agents, it's more of a pain to review someone else's code than it is to get a well-written description of their problem.
>We take contributions as human-written text, not code — see CONTRIBUTING.md. Describe the change you'd like informally in a .txt or .md file in adrs/, and if we're aligned we'll handle the implementation. Report vulnerabilities privately — see SECURITY.md, not a public issue.
Starting to think people were right when they talked about our industry itself having an AI psychosis problem.
As someone who has maintained an open source project, I much prefer written bug reports and feature requests to drive-by PRs. (I almost don't even care if they are LLM-written.)
I appreciate open source maintainers and understand that it's a lot of thankless work, and I know what I'm about to say comes off as (and probably is) ignorant but the hoops projects make me jump through to report hugs or security issues is often like working for corporate in terms of bureaucracy and a lot of times I just don't bother. I've reported a few dozen bugs so I'm not prolific here but also not speaking without any experience at all. And then you often have automation (eg. Debian) closing bugs because nobody looked at it and saying to reopen it if it's still a problem in the current release.
Like I said, I do understand and appreciate how annoying it must be. But there are two ways to look at this, one is that it's free software a bug report is like a support request - and of course nobody should expect free support. The other way to look at it is that by reporting bugs I'm volunteering as QA for the project and the report is beneficial.
I sometimes add a PR with the fix for a bug report I make, but the last few have been ignored in favor of the maintainer's own code. So I think I'll stop doing PRs and just point to code lines instead.
In their case, it's motivated by a desire to keep copyrighted code out of the SQLite implementation, but I'm sure it has a nice benefit of making it so that an extremely widely used project doesn't get drive by, low effort code review requests while still allowing the community to engage.
Asking that "random people on the internet" don't sent code is not altogether a novel, post-AI idea.
I tend to think we have an industry AI mania problem, but I’m not sure I understand what you find psychotic about this. I find it better to get a text suggestion or description of the change, then work on the implementation myself, even without using an agent, instead of reviewing LLM diffs that I know I will want to tweak to my taste
Sounds to me they're asking people to basically at least put the starting stones to something that looks like a specification. Not a bad idea to gate the flurry of feature requests to people who actually can think 10-15 minutes about the feature they're suggesting/asking for.
This is (probably) a natural evolution of Garry's gbrain, a basic RAG for individual use. Gbrain is super useful as a starting point for an individual; this seems like an extension to a company.
I've found gbrain to be very effective with meetily (instead of granola), vikunja for project planning, various mcp for email, etc. It's nice not being locked in to vendors and models...
Cool. This also feels a lot like what copilot is doing. It’s nice that copilot integrates out of the box with teams/outlook/office, so it has great company/work context
Interesting to see they shipped an "anti-slop" taste skill:
> description: Anti-slop frontend skill for landing pages, portfolios, and redesigns. The agent reads the brief, infers the right design direction, and ships interfaces that do not look templated. Real design systems when applicable, audit-first on redesigns, strict pre-flight check.
> - *PREMIUM-CONSUMER PALETTE BAN (mandatory, second-most-recurring AI-tell):* - For premium-consumer briefs (cookware, wellness, artisan, luxury, heritage craft, DTC home goods, etc.)... - Backgrounds: `#f5f1ea`, `#f7f5f1`...
> Landing pages and portfolios are *visual products*. Text-only pages with fake-screenshot divs are slop.
https://github.com/yc-software/qm/blob/7f2c916360f1797a8ff2a...
> Em-dash (—) is COMPLETELY banned. It is the LLM's signature stylistic crutch and it is the #1 visual Tell in production tests. There is no "limited use" allowance, no "natural language frequency" allowance, no "in body copy is fine" allowance. None.
I guess the em dash is really dead.
Probably not, because now you'll have all of the people who are worried about this ensuring their AI generated text never ever uses em-dashes, and then that will be the new tell.
Or we can just be more concerned with whether something is well written and well presented and accept that sometimes that is going to be AI written text.
My personal "tell" is that I use -- instead of em-dash. If you get a message from me with an actual "—" character it's almost certainly AI generated.
I guess that's kind of my duress code :)
> Or we can just be more concerned with whether something is well written and well presented and accept that sometimes that is going to be AI written text.
Slop is slop, it doesn’t matter if it was human or LLM written. If a code comment or Slack message is 200 words long but still makes no fucking sense then it’s slop. A message can be 10 words long and riddled with typos but if the message is understandable it’s not slop.
I've been posting my writing on the internet since 1990, from a Mac where we all learned how to make en and em dashes, umlauts, accents, degree symbols, proper single and double quotes, and so forth from Day #1 and I have absolutely no plans to stop using em dashes.
Nice try, ChatGPT.
Though I agree with GP: models and harnesses are being updated so hard to avoid any use of em-dashes that soon it will be a tell of human writing. The the pendulum will swing back and forth, forever.
I have been seeing people use it more who I am fairly confident are not otherwise using AI to write public-facing communications. It's almost like there's a minor movement to try and reclaim it. Either that or it's somehow getting past my AIdar.
This quote is so egregiously stupid, I'm instantly disinterested in this product now. No reason to assume the rest isn't the same applied cretinism as the quoted sentence.
It's notable that this appears to be used to generate frontends that look more authentic than the typical AI generated site, not as a test for what's acceptable from a human contributor.
I suppose all the exotic unicode characters are dead now. Anything that's too annoying for a human to type but trivial for AI to generate is probably done for. I used to enjoy using those characters. Sigh.
They have never been annoying to type on a Mac.
And yet they use it on the README
All due respect to the YC folks, but having a skill with 22,069 tokens is a major skill issue. Yes, ironic.
My slop control skill is a thousand tokens. Biggest problem I see with the skill is that everything is prompted via negativa. smh. sorry to be judgmental but it's hard to trust a harness that comes shipped with a skill like this.
Gary is no Paul Graham, technically speaking
The YC folks know engineering hype, not engineering. Most of them have been out of the game so long and focused on other things they've forgotten their fundamentals. You can hear this explicitly in shit Gary says in a lot of videos, it's directionally correct but detached from fundamentals.
Doesn’t this just lead to a new “basin of tastelessness” that, sure, looks different from current slop, but is itself just eventually slop all the same?
So just an accelerated version of the standard UI design trends cycle (which is bad).
Yes.
According to the license, the source of it is actually this https://www.tasteskill.dev/
Which, in my opinion, feels like slop.
Also feels copied from impeccable.style (which I think was the original anti-slop skill).
They all use the same pill glowy status light header eyebrow to immediately tell you they have no actual design skills and follow the herd.
Posers took over the industry.
lol this is hilarious - the “make no mistakes” of design
You can’t prompt an agent to have taste.
Why does every online software say they are "multiplayer" now?
And then like 90% of the time there's nobody "there", and it's barely even collaborative.
If this is "multiplayer", does that make Chrome a Massively Multiplayer Online Browser?
Is everything a game?
Lol. That's a good humour.
Can someone give me an example where this is truly and uniquely useful? I've seen so many of these things and I can't tell if part of it is like mostly for enabling nontechnical folks to do more things or if it's some unlock and additive value. At end of day beneath it all things are just prompts, and then you can provide tools and context, but even then that's not even something I've found that useful to keep because context windows are still limited and bias propagations often need to be constantly corrected, especially when digital artifacts are not perfect representations of the real world.
It gives every employee their own personal assistant (AI agent). It learns their preferences, and more importantly, can access whatever they can.
And what aspect of that is novel exactly?
I didn’t imply it was novel.
A video would've been nice to walk through how it's actually used. I don't mean just a walkthrough after a fresh install, but an instance that's actively used.
Love seeing this direction along with Buzz.
The hardest problem in multiplayer agents, at least for us, has not been the agent loop. It is scoping and QM's per-person scopes plus shared rooms is a sane answer for a company-wide assistant.
I build in the adjacent lane, AQ (aq.dev), a multiplayer coding harness where teams run Claude Code and Codex together), so seeing YC ship "a multiplayer agent harness for work" is validating and a little surreal.
Is Hermes the best openclaw like agent as they mention running it before?
Also, what are power uses really using openclaw like systems for?
Still figuring it out, but it's been really convenient to have an always-on agent that has access to internal systems and can be triggered by webhooks. Some examples of what we use it for:
- automatically fixing simple CI failures
- getting production alerts and automatically creating RCAs and a fix PR
- periodically checking slow DB queries and finding ways to speed them up.
- creating charts to answer one-off questions about our data
I've tried using it as an on-the-go coding agent as well, but found I prefer more interactive agents, so I can see what the code looks like.
I think the interesting challenge isn’t running agents, it’s reviewing their work. The more code agents produce, the more important provenance, review ergonomics, and trust become. I also suspect repository platforms will need to evolve there over the next few years.
> periodically checking slow DB queries and finding ways to speed them up.
How does this work in practice?
At least for me, I have a couple dozen years of DB experience but robot, given performance metrics, can get really close to optimal on a tactical level (single query or pattern of queries) but can't yet do the full normalize/denormalize level of improvements without supervision. But really solid if you have one misbehaving query and give it explain analyze access on a read only account
Hermes is huge and packed with features you probably don't need. I prefer smaller one I can extend as necessary, there are so many on github now and it is fun to test them but have been impressed with dirge (https://github.com/dirge-code/dirge) not affiliated.
I have one reading my second tier RSS feeds and newsletters and giving me news/market updates filtered for things important to me
I'm not contradicting this but offering a contrast, I like Hermes because it simultaneously lowers barrier of entry and shows you what possibilities are unlocked by agents. I don't think I would have the time, interest or creativity to jump into the deep end by either extending an existing harness or rolling my own from the start. This also isn't an argument for doing just that, I might do so in the future, but critically only after Hermes has shown me what's possible and my preferences are developed.
I completely agree. I started with aichat[0] before the current agent trend and hit all kinds of bumps implementing agentic loops on my own. Then goose[1] showed me what a whole team working toward the same idea could do right before the official Claude Code harness which had all the bells and whistles. Now I know better what I want and its mostly less ram usage and a small set of primitives.
I still use Claude code (and codex and other big contenders) because they know what they are doing and innovate in ways I don't want to miss. And sometimes they are better at tasks.
[0]: 2023, https://github.com/sigoden/aichat [1]: 2024, https://github.com/aaif-goose/goose
I feel the same but my preference is for Cecli (cecli.dev)
It does what I need it to do and since I invested so much time in setting it up and even contributing to development it is my go to for coding and even managing my VPS as well as business tasks
I had the same line of thought and spent a while with nanoclaw before realizing that adding the features I want back in would have made future updates too painful. I ended up switching to Hermes and aggressively disabling tools/skills and it's been pretty fine so far. I got more use cases set up than I did in my time with nanoclaw.
Totally fair, I still use Claude code for tasks since it is so polished. I think I just dont have that complicated of use cases so in the end performance should reflect my lightweight needs as the priority.
Hermes is what I was using but I still found it annoying I often wanted to operate 1-2 levels deeper.
Yesterday I just decided to try writing my own version (100% just for me, not open source, no monetization plan, incredibly custom) and I've been enjoying working on it so far (I know, I know, it's been a day, honeymoon period and all that).
Part of it is I like building software (even if I'm not writing every line) and part is I like having full control. Turns out (and many people have said this) the basic agent loop really isn't all that special. There are a million levers to pull and what not outside the base loop but that the fun part for me. Trying out different ways to add on to the core concept.
I'm really enjoying being untethered for things like "how will I monetize?" or "how do I make this generic so others can use it?". If I need functionality I just add it in, I don't need to make it infinitely pluggable, etc.
All that said, I'm thankful to things like nanoclaw and then Hermes for exposing me to the core ideas. I just want to put my own spin on it.
I didn't have good experience with Hermes. Vellum.ai was better, but unfortunately it had a bug where opencode go providers failed to work for a time, so I instead starting writing my own.
(Now it is working)
I think Hermes is a kitchen sink of antipatterns and bloat, which is true of the vast majority of these "Claws"
I'm currently using it to help me with my oncall, first responder to our any production alerts. It's not as efficient as coding agent by default, but it's been tremendously helpful to me.
hermes is great to get started with, but it's packed to the gills with stuff you'll probably use one time just to test it. and this eats in to your context so if you're hoping to run it on a lighter-weight local model you'll run in to some trouble. if you go into it planning to customize/thin it out it's solid
I think most people use it to poll their email and instant messages and whateva with an LLM
Honestly the limitations/security of it kind of made it a novelty for me. I use web hosted stuff like surfboard now for my llm-assistant work stuff.
So that explains why there is so much low effort cold outreach on LinkedIn from YC founders these days.
It's getting ridiculous the amount of unsupervised agents doing active inbox management on things that should be personal relationship work. I'm so tired of it.
I gave an agent its own Slack channel and it started scheduling meetings with other agents without me. I've never felt more like middle management
I hope the other agents replied with 'This meeting could have been an email.'
It's fascinating to see new UI primitives and concepts get invented in the LLM era. The sea of creativity makes it hard to even understand most of what each new app does, and nobody describes them well. When I went to the Hermes agent web page, I was left with zero clue about what it did or what it could do. It took a bit of digging to find the right part of the qm page that helped me grok what was going on.
I've become attached to Orca (yc-backed) for managing coding sessions in the past week, but some sort of postgres session db is what's really lacking. So, maybe it's time to try qm.
It is fascinating and I love to see it. Ultimately though, why not build your own? I think that in part is what we're beginning to see-- highly customized and personalized software. I take most of these as inspiration these days and just build my own. Nothing you can't hammer out with a few good Claude sessions.
But why build your own from scratch when you can do it on top one extensible UI platform that already has all of your organizational context, not to mention coding agents already built-in that you can just tell it to build & deploy your personalized workflow softwaree from scratch? ;-)
Because it's fun! I don't "really" use LLM for anything serious yet, but after discovering playing with a custom harness I licked blood. Like going back to edit any message to start a new conversation branch off that, or being able to go back and remove things that turned out to to be dead ends from the history. Not really "useful" [0] for now, I always wondered about that and I was dumbfounded how trivial and tiny a first MVP was. Of course, that's also because it is trivial, it's just chat, it affords none of the usefulness you'd want for any "work", but it's still very interesting.
And for some simple recurring tasks or local housekeeping stuff, having something where I 100% know what files it reads or writes will surely be useful. E.g. if I wanted to watch out for certain topics on HN, I'd rather make something myself that grabs the feed and turns that into a list of titles and topic ID, which is probably 1% of what the front page HTML would be, and then only have the LLM process that output -- rather than telling an LLM to do all that every time. Even if tokens may not be that precious, and the difference in "cognitive performance" not worth speaking of, that would feel way neater to me.
[0] https://imgur.com/a/ZL8dhYg
That's an option right? When I say roll your own-- it could be extending another platform. There's definitely different levels. Customization is in reach regardless of where you want to start and stop your stack.
Why not build my own? Because my good ideas are in other areas, and I'm going to expend my energies where they have the most impact.
I've replaced most of my business software that I was paying for with custom stuff, already, for busy-LLM-work.
I am but a leaf riding on top of the sea of creativity of others when it comes to these new interaction patterns.
These new experiences are built atop layers and layers of bedrock libraries that we, generally, don't reinvent. Or we reinvent one or two.
What's notable to me is that ux doesn't so far generally have this behavior. That as per this post people just build a new app, a new experience.
I want to believe over time we'll have better composable & malleable ux experiences atop broader platforms for us. That over time the "go it alone" path has other worth ways to innovate that use a more substantial shared base. It's dangerous to go it alone, and doing so equipped with just our wooden sword and some courage and perhaps an LLM wisp is an amazing adventure, but I think the survival rate & impact would be much better if we had more general ux systems that supported better innovation atop them, and if less people did the pure "why not build your own" path.
I understand you take. And I definitely agree that we could use a lot better ux tools than we have right now. Maybe one of these prototypes will lead to that though. I'll look at it more as creative engineering. Not everything is meant to be distributed or consumed. But that doesn't mean that it's not worthwhile to try and prototype different flows and options. I have seen hundreds of harnesses, I've only used a couple-- but they've inspired me to build something that fits me better and works for me. If somebody comes out with something that helps unify the UX experience and provides better building blocks for me to work with, I'll definitely switch to that and rebuild my experience on top of that. But for now this is what we have. Admittedly this reflects my level of experience and my expertise too. I don't have a lot of experience working with building UX frameworks, so I adapt what I can.
People are using the UX systems though, just that it's the agent that is gluing them together now. Under the hood, it's still React plus component libraries like shadcn, lucide, tailwind, axios and so on.
The agents are so good at this, there's literally no point in not doing it that way.
These are all surface veneer but so so so much less than an app. The entire rest of the owl has to be redrawn, and all the guts inside the owl redesigned.
UX does have this, and it has for many years. You might remember Bootstrap way back in the day, and everything looking like bootstrap. These days it seems like everything is based on tailwind under the hood.
Right now shadcn/ui might be the leading candidate but there are plenty including old school bootstrap, material / MUI, a bunch of various flat-themed ui, etc. They often come in the form of a react component library or whatever but they're generally really solid. I also found one in svelte that I can't immediately pull up but it was nice looking too.
It's really hard to focus on the unique aspect of your tool because it runs the risk that people will think it's unfamiliar and therefore not useful to them. That's why a lot of the tool marketing pages look the same even though the tools themselves might be different or innovative
AI needs entirely new primitives in many areas.
I think we need a new area of study around UI/Agent connection. It can kinda be done with tools, but I'd we need much deeper primitives to allow the UI to inform the Agent and vice-versa. Right now we've just given up, replacing the UI with an Agent message view, but I think that's just because nobody is thinking about how the two can compliment each other
I've been playing around with using a hidden markov model informed by a UI state event stream, with the end state fed into the Agent as a hint on each message turn. Then the Agent can make a tool call to add events to the HMM. This has been really interesting, but I haven't struck the right balance to make it actually feel good for the user yet
Isn't this the direction Claude on the web is going towards? It's a mix of message view on the left and specific interactive UI on the right panel. There is also MCP Apps spec now
It's hard to say what is needed now, vs needed in a few years.
For example, many rumors say SSI has solved learning and retaining state. That would significantly change AI requirements.
Aren't there a ton of products already doing this? Why not just use claude Cowork? Surely they're simpler/better/more featureful/developed than the alternatives here? What advantage does this have? Would love to see a 'QM vs Cowork' comparison!
> Why not just use claude Cowork?
Because people want to be able to do things like use their own clients of pi or opencode with LLMs they run themselves, such as the just released deepseek v4 flash 0731, not permanently tied to an Anthropic ecosystem of non-open-weight LLMs and pay forever per token.
> Would love to see a 'QM vs Cowork' comparison!
It feels like the big thing they are touting here is the shared company brain. Not clear to me though, how that brain is developed when each person has their own harness. (I did only skim the docs.)
And it still has the issue of: if the agent is acting as me, then security wise it can do anything I can do. Maybe that's why they are recommending for startups.
But yes, comparison would be helpful.
> Why not just use claude Cowork?
Many people don't want to support a company pushing for regulatory capture.
Check the speakers at the startup school that happened just last weekend…
this one is trying to cashin on the 'multiplayer' keyword after the buzz that 'no one has cracked multiplayer yet'. well the ui sucks and is not it.
Kinda makes sense. Multiplayer is at odds with async, generally speaking - and LLMs/agents really see their productivity rise up with async work.
I disagree with this. If agents are doing asynchronous work, it's really an artifact anyone should be able to look at.
what does multiplayer means in this context?
People are building state-of-the-art harnesses in private right now that outperform the average developers' harnesses, people using Claude Code, OpenCode, etc.
maybe you want to use a different model?
I love the concept to use different harness frameworks in qm but a true multiplayer harness needs to support other agents and any MCP clients, including Cowork.
Making agents multiplayer is mostly a context problem. You could be using ChatGPT or a Slack bot, or a web interface and the agent needs to know you, your conversations in Cowork etc. so it can enable multi channel collaboration with your agents and your colleagues. We're working on it at https://lobu.ai
Pretty cool - I’ve also been doing something along the lines of this with lumifyhub but it includes docs and boards natively
This was literally in YC's Request For Startups for Fall 2026: https://www.ycombinator.com/rfs#multiplayer-ai
Multiplayer agent harness is useful only when the context and AI output does not overwhelm the team space and works asynchronously and background. If you think about it, that could be just boring job scheduler, isn't it?
What is yc software?
A team of brilliant people who work at YC shipping products internally and externally.
>brilliant
A team that pioneered writing small checks for outsized returns. AI is making that venture model a lot less attractive, so now they're building productivity software for the startups they once only invested in.
I think it's the website you're on
Someone from Y Combinator made some software. Must there be narrativization of everything?
Yes, please!
I'd like to know more, for example why now?
At least for me, I often have a hard time believing people’s stated reasons for whatever they’re doing.
The substance <> narrative relationship is backwards a lot of the time. Someone does something for a nebulous multitude of reasons, and then post-hoc fits their decision-making into a logical explanation that sounds nice.
There is some psychology research supporting this as well.
I suppose if you see the narrative itself as part of the release, that might be interesting. But most of the time I’d rather just hear plainly and straightforwardly what the thing is.
Or at the very least, I am very accepting of releases which do not include rationalization / narrativization and don’t think it’s required to include.
qm has been a qemu (virtualization) command line tool probably for 20 years or so. The ignorance of these people inside the same field is egregious. (Oh, yes, there can be tools that are named the same in the same field, I get it)
In a tangential earlier release, Gary Tan's own gstack:
https://github.com/garrytan/gstack
Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA
YC CEO says he ships 37K LoC AI code per day. A developer looked under the hood https://news.ycombinator.com/item?id=48815117 (July 5, 2026)
This is the exact same as the breathless faddishness of celebrity news. "Use these seventeen different skincare products and you'll look exactly like Gwyneth Paltrow".
> We take contributions as human-written text, not code — see CONTRIBUTING.md. Describe the change you'd like informally in a .txt or .md file in adrs/, and if we're aligned we'll handle the implementation.
Interesting approach to open source contributions. Closer to feature requests at that point?
I'll need to explore how they're doing org wide context and security for sure. This seems extremely complementary to my own coding tool which currently gives the best AI interface for individuals.
Would be cool if in a few epic tickets i'll have both their org wide architecture AND a productive individual coding interface :)
If I had applied to yc, I would have thought they had "distilled" my startup idea xD
It's not very novel
Ok but then why did they ask for this in a Request For Startups?
If they asked for it, that's a big clue it's not a novel idea.
They ask for it because they think it can unicorn. Being novel isn't always important.
If it is so not novel, why did I never see it in Enterprise? Why is their own attempt a barebone mvp? Besides Dassault Systemes or Palantir with their Ontology product, who was doing anything adjacent? It's giving typical HN vibes ' you don't need dropbox, just sync to ftp' ;D
i'm curious what this gains in practice and usability over standard chatbot slack integration.
looks like an internal tool that yc rushed out the door to minimize any most lost ground to Buzz. That said, I'd be curious what folks think comparing these two tools.
Fake, boring, non-accomplishment.
it says "Each deployment runs in the operator's own cloud account"
But feels like its written to run on one mac/vm and carries same drawbacks of other similar platforms. I'd rather use Hermes/Openclaw for oss or closed managed agents like Tasklet or Prajvis
Oh sweet, nice to see someone writing about multivalue databases.
The design concepts of “qm” and “Block Buzz” share some similarities in their underlying principles.
Just want to thank the authors for the concise and consumable README.
> Learn your writing voice from past sends, then triage your inbox on a schedule — labels and reply drafts included
Oh grand. The world will be full of agents talking to agents and nothing getting done. So pretty much the same only I can just leave my machine awake to look busy.
i find something a bit funny in an ai project, written by ai, requiring human-written text with specific guidance to not use ai.
"Given that coding agents write most underlying code now, we'd prefer PRs in the form of human-written text. [...] Please do not have AI artificially expand what you'd like to do into a formal proposal."
XY problem, don't expand the design doc before you have the idea for the design. Don't compile down into lower abstraction levels until necessary
i think it makes perfect sense and I'll probably adopt the same posture for my open source project.
The idea is that the core dev team is the one with the AI harness. If someone contributes an idea they can just feed it to their AI to implement it and it probably costs 10 minutes of human time, because they trust that their harness will be able to implement the feature competently.
So, because the implementation effort is so low, the only aspect that matters if the quality of the idea. It's way easier to screen human written text than a bunch of code in a PR. If you just give them code, they would not know if you made it with a competent harness. If you gave them a AI written design they would have a lot more to read through to decide whether it's slop. If you just give them an idea it's a lot easier to determine whether it's high quality.
Yet the README is generated ("Two skills maintain the boundary in both directions.") and the demo is .. a Mobius strip?
I'd prefer if you explain what it is you are building in the form of human-written text.
Well, if they're consistent, the ADR directory contains only human written text that drives the underlying development... Aaand it's empty. Well.
I personally would probably have the readme generated based on that directory as the primary document, probably with another `readme_generation_rules.md` in the ADRs directory, and I would be pretty ruthless about disallowing all the slop-adjacent wording.
> I would be pretty ruthless about disallowing all the slop-adjacent wording.
I’d think that trying to play whack-a-mole with slop like that would result in a lengthy prompt, as well as being brittle to future changes.
If you want the prompt to be the source of truth, you’re probably going to have to accept prose that feels like AI, at least until you can regenerate with some future smarter model.
Not even close :) It's really, really good at making natural-sounding prose if you demand it, set clear guidelines (ideally a script it can run with grading, but...), Fable in particular is eerie about mimicking prose style. Try RFC style if you want really clear architecture documents with MUST/SHOULD etc.
It will, of course, be slightly worse at the non-style task (i.e. make the readme worse), but if you care about the style, that's probably an acceptable tradeoff.
Coding agents are extremely useful but often extremely dumb with design. If you do not design the software yourself, you will probably get slop. This was always the core issue with vibe coding.
AI Averages, and is inclined to do average designs and implementations, which in some cases might be an improvement, but long term it creates more to deal with.
Hm. If "AI averages," then why doesn't it create an average amount to deal with, instead of more?
Because competent people are evaluating it. In places where the standards are rock-bottom, the AI is leagues ahead of anything they'd produce normally.
Because the “average” developer is incompetent.
They specifically ask not to create/submit a formal proposal. I do not see any restriction on the use of AI otherwise.
i read the emphasized "human-written" part, in combination with that last line, as a blanket restriction (for contributing). but perhaps i read it wrong.
List of commits already include at least two contributors who are openly working with Claude on their code: https://github.com/yc-software/qm/commits/main/
The point is that they just want your prompts, not your code.
yes, specifically they want purely human-written ones.
i understand why they want their own agent to do the code, and i can see the reasoning behind it. but not allowing ai to format/tidy up/flesh out the proposal is the part that i thought was a little funny.
It's definitely a little bit funny. I think it's a good and interesting rule though!
That's not surprising. I have much the same thing in my latest open-source work[0]. I have a workflow that's working nicely with agents, it's more of a pain to review someone else's code than it is to get a well-written description of their problem.
0: https://github.com/pjlsergeant/byre?tab=contributing-ov-file
not a very helpful title? maybe: "qm - a multiplayer agent harness for work"
It's an HN-ism that it's OK for titles to make readers do a little work.
>We take contributions as human-written text, not code — see CONTRIBUTING.md. Describe the change you'd like informally in a .txt or .md file in adrs/, and if we're aligned we'll handle the implementation. Report vulnerabilities privately — see SECURITY.md, not a public issue.
Starting to think people were right when they talked about our industry itself having an AI psychosis problem.
As someone who has maintained an open source project, I much prefer written bug reports and feature requests to drive-by PRs. (I almost don't even care if they are LLM-written.)
I appreciate open source maintainers and understand that it's a lot of thankless work, and I know what I'm about to say comes off as (and probably is) ignorant but the hoops projects make me jump through to report hugs or security issues is often like working for corporate in terms of bureaucracy and a lot of times I just don't bother. I've reported a few dozen bugs so I'm not prolific here but also not speaking without any experience at all. And then you often have automation (eg. Debian) closing bugs because nobody looked at it and saying to reopen it if it's still a problem in the current release.
Like I said, I do understand and appreciate how annoying it must be. But there are two ways to look at this, one is that it's free software a bug report is like a support request - and of course nobody should expect free support. The other way to look at it is that by reporting bugs I'm volunteering as QA for the project and the report is beneficial.
I sometimes add a PR with the fix for a bug report I make, but the last few have been ignored in favor of the maintainer's own code. So I think I'll stop doing PRs and just point to code lines instead.
SQLite has a conceptually similar contribution process:
> the project does not accept patches from random people on the internet
https://sqlite.org/copyright.html
In their case, it's motivated by a desire to keep copyrighted code out of the SQLite implementation, but I'm sure it has a nice benefit of making it so that an extremely widely used project doesn't get drive by, low effort code review requests while still allowing the community to engage.
Asking that "random people on the internet" don't sent code is not altogether a novel, post-AI idea.
I tend to think we have an industry AI mania problem, but I’m not sure I understand what you find psychotic about this. I find it better to get a text suggestion or description of the change, then work on the implementation myself, even without using an agent, instead of reviewing LLM diffs that I know I will want to tweak to my taste
Sounds to me they're asking people to basically at least put the starting stones to something that looks like a specification. Not a bad idea to gate the flurry of feature requests to people who actually can think 10-15 minutes about the feature they're suggesting/asking for.
as a maintainer, seems nicer than getting a slop pr with no context.
the other option is to generate 5000+ open issues that no one cares are open.
Also, it's a AI project; what, exactly, do you think they're going to try and do?
So, basically,
> Please write our prompts for us
This is smelling like an astroturfing campaign. This is a total nothing-burger
Sweet, now YC itself will be the startup :D.
"multiplayer"? Is this a game? I honestly don't know what that means in this context.
this looks over engineered, gimmicky and lame. YC is clearly having a "crisis of meaning" in a post AI world
This is (probably) a natural evolution of Garry's gbrain, a basic RAG for individual use. Gbrain is super useful as a starting point for an individual; this seems like an extension to a company.
I've found gbrain to be very effective with meetily (instead of granola), vikunja for project planning, various mcp for email, etc. It's nice not being locked in to vendors and models...
Thanks for making this OSS too.
Heavy conclusion from a single OSS release
vibes.
Other than the hosting providers, who has made money directly from running OpenClaw in constant loops?
This software appears to be yet another solution in search of a problem designed to burn as many tokens as possible.
Real and useful, thanks
Cool. This also feels a lot like what copilot is doing. It’s nice that copilot integrates out of the box with teams/outlook/office, so it has great company/work context