Love to see these crazy kinds of experiments going on. Even if this doesn't 100% work or is prohibitively expensive for now, these are glimpses into the future in the same way people were talking about coding agents in 2023 when we just had tab complete.
It shouldn't be prohibitively expensive because the more you break the work into smaller parts, theoretically the smaller a model can be that completes it.
Therefore your agent swarms don't have to be cutting edge API calls, they can be hosted on clusters of tiny cheap computers with small CPU GPU. For orgs you could basically invest in a local server that has 100s of small computers.
You can use engineers and frontier models to plan and do task subdivision. And then shit that out to the local cluster that's basically handling each small task without knowing what it's supposed to do.
The real issue is the middle. Testing, verifying interconnectivity of mid level abstractions, building that amount of tech debt at such a high rate and actually being aware in any way what it is that you've built.
I think for most MBAs, considering the US and a significant amount of global markets only care about short term market upside, and AI is still in the middle of the hype cycle, that's not a serious issue for anyone who is profit oriented.
It's only a real issue for the losers like you and me who care about sustainability and infrastructure that survives longer than one market hurricane season.
For my current setup, the most efficient way is to use larger models for coordination, but a heretic'ed qwen 30B model for the implementations. If you build your agentic environment around specifications and test coverage tied to symbols, you can do a lot of parallelization of agent work.
If you then separate the filesystem and tool read/write access by agent roles and policies (e.g. coder not allowed to modify unit tests, tester not allowed to modify code files) then you can force them to use a centralized per symbol/per contract specification tool.
And then you can just let agents discuss issues, where the messaging threads are also tied to the same symbols.
The twin trends of AI slop being churned out at tremendous rate and deployed in everything, plus frontier AIs becoming really good at exploiting many small security holes to pwn boxen, is going to collide in an amazing cyberstorm that was predicted by every cyberpunk novel from the 80s X-D and we are gonna have front row seats 8-D
I am afraid the folks are still busy playing with familiar toys. "Let's extract more code out of LLM's inner layers!" SQLite we already have. That's where their "spec" came from.
The interesting frontier is elsewhere: how to fit it into the pipeline to build new useful things.
I felt motivated enough to write a follow-up/reply post http://replicated.live/blog/follow-up
It's cool stuff, but I think mostly in an aspirational way right now. The tired cliche hit on the tired cliche claim will be "but sqlite is in distribution" which will miss the point, that objection dissolves at the precise point where it would have operationalized: novel software of value. Is SQLite in the train set, yeah. Does that matter? Only if someone is willing to pay for Rust SQLite as something other than marketing.
It's cool but I'd rather know what their most successful users are doing, and how they measure that.
>The browser swarm from earlier this year peaked at roughly 1,000 commits per hour on Git. The new system peaks at around 1,000 commits per second.
>To facilitate this rate of activity, we built a new version control system (VCS) from scratch. Throughput was not the only reason to own this layer. Every change in the system passes through the VCS, so it is where collisions first become visible, and several of the coordination mechanisms in the next section are implemented directly inside of it.
Talk about inventing the universe to make a button.
Making a high-throughput revision control system is doable, there have been experiments here on HN with git-in-postgres etc. The question is, what is the actual bottleneck here?
I made a VCS for LLM assisted work, bootstrapping it on its own code as it was developed. In fact, if it did not really work, it would not even exist!
In my experience, keeping track of things and evaluating them was the bottleneck and I focused on that https://replicated.live/blog/worktree Effectively, the question is "how to leverage human evaluation and feedback".
The Cursor people concluded their last post with "taste, judgement, and direction came from humans". Sounded surprisingly sane after "don't compile it, commit it faster" some paragraphs earlier.
People often think of tech/business as bottleneck issues ie whats my limiting factor. That's relevant to pre-ai business as humans are relatively self organizing.
The issue with AI is it has no real memory, context, long term reasoning, etc. These are a lot of what we use to determine how well we're doing something as well as why.
The "bottleneck" so to speak with ai is that it will just circularly build junk forever, expand scope infinitely, and just continue building.
At its best it will autobuild your unicorn saas. But on average it's a cancer cluster of ever expanding code with no plan.
The issue with unsupervised AI is that when you increase the throughput you decrease the amount of human supervision. So you have to have some way to deal with that.
If I'm going to give AI a gigantic task that's really complicated and it will probably mess it up 100s of times, how will I compensate for that without constantly looking over it's shoulder?
Some critical parts of agentic development are literally months old while most mature ones are merely a couple years old.
Critical thinking was the most critical checkpoint for models to reach.
Now its only a matter of time for the right tooling and orchestration to come along. Once agent memory and recall works well, all of those bottlenecks will go away one by one.
Honestly judging from the volume of his output, a ton.
It’s rare to run into prodigious artists who simply produce a few perfect pieces - as a rule those few magnum opuses are built on a hill of cutting room floor scraps.
The director James Cameron stands out because he has only produced huge box office hits. Extremely hard to do like you said, and most artists and auteurs would be happy to have one great work they can be remembered for, even if it's amidst a dozen forgettable pieces.
I don’t know how to feel about HN thinking one of the greatest artist of all time produced slop. You guys really think Leonardo davinci was drawing ladies with 6 fingers before the Mona Lisa?
I assure the worst of Shakespeares work is far beyond anything most writers can produce, much less the incoherent slop LLMs generate
The problem is the infinite other literature produced along the way. It doesn't work to throw more monkeys at the problem of verification, and checking for matches against existing Shakespeare plays is cheating.
Right. The joke with shakespeares monkeys or borjes library is that you just do that and then filter out the books that match shakespeares works.
If you already know what you want in an infinite library, it is there. If you don't know exactly what you're looking for to the letter, you might not be able to find it.
Yes, because the folks waiting slack-jawed for the monkeys to finish will no longer have the skills or discernment to know what Shakespeare looks like.
And if in the end it does not produce masterworks, going faster and breaking things will be seen as a huge and credulous waste of time that poisoned the internet and some codebases beyond redemption.
But I’m glad someone else is running all these experiments so we can see who is right in their guess about the outcome.
2*128 is the UUID namespace, which is astronomical. That is like 21 letters. Making one sentence of Shakespeare by random search is strictly 0% chance. SHA256 is ~43 letters then. "LLM broke all of cryptography and creates random data to fit a hash" -- are we there yet?
Random search does not work at this scale. In all their cases, they played with pre-existing well-studied software, so they had an embedded fitness function. An extremely detailed one.
Monkeys have been reading Shakespeare all their lives (training). Effectively, this is LLM uncompression. Hence, "in Rust, from nothing but its documentation" -- not entirely true.
Yes it will be a terrible waste of time since dedicated humans can do it faster and produce better stuff. I would prefer literature created like that instead of just random monkeys writing what they don't understand. We also lose the ability for writers and creators to get better at their craft and produce unique works instead of just randoming spamming keys for millions of years
Isn't that the core fundamental of AI? Constrain the results to a latent space which has all the probable answers and then pick up whichever token makes sense based on your random sampling?
I feel like the whole concept of harness and LLMs IS to pick random tokens which make sense. Think about this as a solution space (a very complex one), then the harness and the LLM are moving towards the local minima at which the code is complete. This is akin to how we think about AI anyways.
Yes, agentic coding is machine learning. Not because of the underlying language models, but because it is gradient descent, where the test harness provides a gradient for the agent to traverse.
> To test that progress, we returned to a task the old swarm had struggled with: building SQLite from scratch, in Rust, from nothing but its documentation.
Even if no Rust code for it was seen during training, an LLM can trivially transpile SQLite's C codebase to Rust on the fly. For example, I just asked ChatGPT to write John Carmack's famous Fast Inverse Square Root algorithm in Erlang, without searching online or thinking, and it transpiled it immediately (while also extracting the knowledge in the same step). SQLite's semantics/code are stored in the middle layers of an LLM, and the last layers are able to convert it into any representation, as conditioned by the prompt. Cursor's experiment is deeply flawed because they merely extracted the model's compressed, lossy knowledge of SQLite's codebase and then just ran a bunch of tests/fixing rounds to make up for the lossiness. The claim that the agents built it from scratch is false.
Then you would expect the implementation to be structured the same as SQLite, and having glanced at the result, it looks like at least some things aren't. For example, it seems to use an operator-tree executor rather than SQLite's bytecode interpreter.
The point is that the model has the source code in its weights. Therefore it's misleading to focus on being "only given the documentation", when it could produce a SQLite implementation in rust with no documentation at all. The wording is designed to create a marketing impression that it can do spec-driven development for a novel project, but this example doesn't actually support that.
It's like a diet system advertising how much weight a spokesperson lost on their system, without disclosing that they also went on Ozempic at the same time.
The knowledge is lossy, and code generation itself is non-deterministic (temperature), so the operator-tree executor must be interference from other DB implementations, because it's uncommon to have a bytecode interpreter
Why tho? The whole point of agentic is that it takes several passes to get to a result. If step a) is a non-idiomatic transpile, but step b) c) and ... z) make it idiomatic and performant and efficient, why look at step a)? That was just the beginning. The idea with "agents" and "swarms" and dynamic this and that is that you only look at the last step.
This strictly follows the early agentic usage, where people were complaining of code full of errors and such. But having the loop do plan-write-verify is the thing that got us from that state to today. And it works. If you look at the code and it has errors (linting, etc) then your loop is not properly setup. You should only be called upon when all that can and should pass does. Your role is "direction/taste/human-common-sense-at-a-glance".
Because the point of the test is to see what the agentic swarm approach is capable of. If your test assumes that a version of the thing you are trying to build is available in the training data, or can't determine if it matters whether the thing you're building is in the training data, then it's not a very useful test. Most of the time when you consider agentic approaches, you're not trying to rebuild something the agents involved are trained on.
How do we know if these models weren’t trained on Turso’s rewrite of SQLite in Rust?
It seems both likely that they were and impossible to remove that code from pretraining. Doesn’t that make this just about LLM memorization of the training set? What am I missing?
I call this junk "meta-agentic engineering", it reminds me of people who have the coolest nvim configs, spend hundreds of hours customizing it but ultimately get less work done than the guy with minimal workflows, if any at all.
I look on twitter and it's just people building tools for agents to use agents, some weird customization loop going on in the LLM space right now. Ultimately these are trends pushed on us from model providers because they 10x token consumption. Its literally just BS trends to increase revenue at these companies, most of it is largely useless.
Some of this stuff is ludicrous. I finally tried /loop last week and discovered every loop iteration passes the entire context history to the model. So pretty quickly you're running a full 1M context window, without cache, likely just to check if something is ready or needs to be done. It's miserably terrible engineering unless your entire and only goal is to burn tokens.
"On a Claude subscription, Claude Code requests the one-hour TTL automatically. Usage is included in your plan rather than billed per token, so the longer TTL costs you nothing extra and only affects how long your cache stays warm.
If you’ve gone over your plan’s usage limit and Claude Code is drawing on usage credits, you are billed for that usage, so Claude Code automatically drops the TTL to five minutes."
I never liked the term "ricing" with window managers, but this is definitely veering into LLM ricing territory sometimes. And people are getting rich off of it.
I think many of us engineers (including myself) are suffering from AI-induced procrastination. Fixing that pesky bug in production isn't fun. Building a harness/workflow that detects and fixes pesky bugs in production on its own is way more fun (especially when your employer pays for tokens), and it feels just as important, if not more! So what if it didn't actually fix that pesky bug in production on its own...
I'd rather anyone using llms to fix pesky bugs in production gets fired and I work at a shop that cares about quality and craftsmanship. There are handmade swiss watchmakers, and then there are assembly line workers in China also making watches, which worker do you think goes home happier at night (and who's company is actually beloved), you have a choice.
I don't think there is a lot test harness, given there is a golden implementation (sqlite3), with clear interface (SQL, and sqlite3 db files). Once you slop crappy implementation, then you can create one group of agents finding gap where slop doing wrong vs golden, and another group of agents filling the gap meeting the conformance.
I assume they did similar, as long as test case(or spec) can be parallelized. Having parallel execution / testing environment is crucial if we need 1000 commits/second throughput.
I would not surprised they have 1+ millions of specs.
I would bet they won't release generated specs - actually this will be a valuable asset like https://sqlite.org/th3.html
Yea, looks like it (at least for their v2). Fable did much better in V1, but once they added their tooling around it, Opus + Composer ended up doing better (>0.8 Grade on the SQLite chart) for the final product. Granted, Fable reached a 'good' result (0.7~) in a much faster timeframe.
Point is, frontier models don't necessarily are better than a swarm of non-frontier ones for these scoped problems. IMO frotnier models excel at underspec'd or more complex problems where ideation and exploration are key (and that's where they become crazy expensive).
Anthropic wants it that way so they can’t be sued… the correct answer is don’t, switch to Kimi K3 for much more usage, the same quality of model, and no hidden practices :/
I'm not sure of the long-term viability of these brute force swarm approaches, but clearly the AI companies dig the marketing they get out of it. But the fact that they get so far at this stage suggests we have a very interesting 2 years coming up.
This is almost a year behind Steve Yegge's first post on beads. Gas Town and Gas City provide orchestration for the swarm. So far I haven't seen a perfect implementation but this idea isn't new.
I don't think Gas Town was actually all that much ahead of the curve - see Cursor's swarm post from January this year, mentioned at the very top of this blog post.
I also don't know that beads is a very strong example that Yegge knows what he's doing - it's famously been described as (pseudo)malware.
Everything Steve yegge has done has been trash. That's why nobody is talking about beads or gas town. It was clear even in the beginning that it was a borderline AI-psychosis-fueled trash fire.
I think if you're suggesting that Steve Yegge somehow invented having agents collaborate, you're way off base. People started talking about agents collaborating very, very early on, and people have been doing experiments like this for ages. There have been libraries to build e.g. graph-based (possibly multi-agent) workflows from the early days, before even structured output was a standard thing in the APIs.
The only thing Yegge did was come up with really stupid, convoluted and anthropomorphized language to describe the process and then write unhinged articles about it
Even at a much smaller scale of a few dozen parallel agents working I've found real benefit to a structured hierarchy of agents. Not just keeping the context clean and focused at the right level, but if you find that something has just gone off the rails, you can more easily rip out or even just fully delete a leg of work without damaging the original design. It's also nice when you find a design change is needed and you can communicate that to the top level and have it efficiently propagate down where it matters.
I like it so much I even recorded a video walking through the process!
I initially thought that getting agents to work for longer and in large groups was the future, but I'm increasingly thinking that, at least for engineering, just one thread makes more sense. The agent pulls things into context as needed. One thing that I've been experimenting with is also letting the agent remove things from context, such as files. But just adding to the context and compacting when it's full seems like it might beat a lot of more advanced options. Because the model is good, it knows what to put in the summary; just enough for the model to be able to rebuild the context from that seed (e.g. pulling in relevant files/data into the context).
Have you used this method yourself for long workflows and with contexts approaching 1 million tokens? It doesn't work very well. LLM context is nearly half unusable.
I always let GPT 5.5/5.6 sol reach compatction. I'm only offered a 258K window, but I think once they release 1M to normal plan users, it will be usable across the whole 1M. At least with Opus I could use the whole 1M. I'd argue that the more context I use, the better performance gets. It has more info already available. I don't notice degradation.
This is super fascinating, and I loved seeing testing of where exactly you need frontier intelligence -- looks like coordination / planning, but not coding right now -- the article's a bit of a tease, as we can't play with such a harness, or their new version control system or get a workable artifact out of it.
That said, I love the work on figuring out these harness coordination jobs. While there are analogs to human management there is also this enticing feeling that, since the models are broadly deterministic, we might be able to get repeatable science-type lessons about managing them with enough testing.
Cursor is saying we need to get improve the harness and tooling, but the models are getting better. One would have to assume that will continue and that cheaper/smaller models will continue to become even more useful.
These tests are interesting but I can't help feeling that these are benchmaxxing results. "We made our harness make sqlite in rust" sounds like a much easier probelm statement than having to recreate an application like facebook. I think the concept of having to integrate is the real challenge. Re-writing code in new languages based on decent docs is a difficult but less useful measure of how AI can help replace engineers.
Imagine this: you need to build a new CRUD app but your AI agent spends 20M toks on just writing it's own custom login function, that's a waste of time and money. I would not like to have a new system developed for me in such cases. I think we need a test where the agents are more specifically tasked to find technologies worth integrating and hence write the fewest lines of code.
The nature of ML is to reach for easily measured proxy problems which can then be optimized. Why worry about a backoffice integrated micro-service you have no data on when you aren't sure that you can design and implement a rewrite of an OSS project?
This post describes ideas found in existing agent orchestrator systems since they've existed. You'll find the same "top level agent that breaks a problem down into spec / plan / implement / verify" pattern in get-shit-done or superpowers.
Is the original part the VCS that works at 1000 commits/second? Or spending a bunch of money comparing models?
If you can't see cases for it you may need some glasses. Replace the graph in the docs of Top-level planner -> mid-level planners -> worker leaf nodes and you get an org chart of CEO -> executives, N-levels of management, and ICs at the leafs. It directly maps to almost any office-based company that can operate completely digitally.
It's obviously in its infancy but fully agentic companies are becoming a possibility, given the right harness, models and tools
Lets separate "workers" from "decision seats", you might need a lot of workers which typically don't actual require inference and might actually just be cheap solvers but actually less decision seats. So i guess it depends on what you are calling agents in your scenario.
I find these blog posts (and the originals, with Anthropic's C compiler and Cursor's browser) somewhat funny, as if they have this enormous power to build ... but they can't build something unique or new. Like the software sucks, but look how powerful the process is (the models are indeed powerful).
And it's a bit of a shame: by virtue of their position (their embedding in the fabric of venture capitalism), it seems like they can only make a subset of things -- what they can make is dictated enormously by capital, as they are engines of capital.
Not sure the point I'm trying to make, I just find it amusing.
Perhaps the point is that it might be more worthwhile to give independent creators a billion dollars to play around with agent swarms if we want to keep diversity in the evolutionary algorithm that is the software industry high.
Most of these reusable, automated workflows sound really cool in theory, but in practice either they require a lot of customization (not really reusable) or a lot of babysitting (not really automated).
AI has improved my productivity significantly, I'm not a hater, but we also have to be honest about its (current) limitations. The whole "software factory" thing is just a pipe dream people sell to execs / investors / online courses.
Yes and this can be proven because what Anthropic says and does is vastly different
They say SaaS is dead and you can vibe code any SaaS using Opus and yet despite having Mythos with them they have to use JS terminal to run Claude Code.
Why don't they just vibe code CC? Why doesn't Dario have a 100000 agent swarm to run entirety of Anthropic or at least software department?
My app is written in TypeScript. With agents I could rewrite it in Rust, I just don’t need to so I don’t. It doesn’t warrant that level of performance and the business logic is more digestible in its current form. Given the amount of reviewing I need to do now, readability is worth more than performance. There are always tradeoffs.
Are you implying that them attempting a Rust rewrite would fail, and this is why they’re not doing it?
I do somewhat agree. I can only speculate that they're pumping out new features incredibly fast and they don't want to fall a month behind the competition.
What do you mean by “they spent $ telling everyone that”? If anything, they told everyone that to receive $ in investments. Those statements are pure PR hype.
Also famously terrible. If not for the walled garden and the fact that anthropic can take away our ability to use custom harnesses any day, I would not be using CC. Given some of the features they add I've also started to suspect that their models are making product decisions. Not necessarily intentionally, but more in the sense that I think their engineers just let their models build and add whatever it feels like is a good idea to the codebase and then they just publish that.
Yes, that's true, but in the post, they even mention that building the spec is the scarce resource.
For that to work, the swarm has to actually follow the spec, which is what much of this post is about. We gave the swarm 835 pages of prose and it came back with a database. What was scarce in this experiment, and what we expect to be scarce in software engineering going forward, is the right description of intent.
We're rediscovering the "architect" job role a decade or so after switching to agile processes and staff engineers largely replaced it. God help us when they rediscover UML for agents.
Easily the biggest delta between a good one and a bad one I've ever encountered. It's so easy to test too. Just ask a product manager if they can give a talk on <some product they own> in 2 hours. If they respond by slacking all of the engineers that work under them to stop what they're doing and give them the info to give that talk, fire them. If they are ready to talk right now then give them a raise.
I mean, how long are these pages? Assuming 40 or so lines per page, that would be 33400 lines of spec turned into 200,000 lines of Rust, or about 6 lines of code per line of spec. That's a very detailed spec, much more than any software specification I have ever written in my career. It's not obvious to me that writing this spec would meaningfully save work over incrementally building individual features, especially when you think about what it would take for human beings to meaningfully review whether those 835 pages contain the right intent.
not only that, but what all of this misses is the fact that the specification that they used was a product of writing the software, not the other way round. It would be much, much more difficult to first write a correct specification than to first write a correct program, because programs are formal systems that can be executed to observe their correctness. You can’t do that with a specification.
Well, the tool works well. I am grateful that Nature has seen fit that the same kind of person who can make it doesn’t also have the ability to identify what needs exist in the market because then all software would flow out of OpenAI and Anthropic and the rest of us would run about their legs and peep about. Whether product development will survive the next generation of models is still unknown.
it feels kind of like a concept car a large manufacturer would show off at a motor show. Everyone knows there is no intention to actually ship that thing, but they're exploring concepts and technology and trying to make inroads in what comes next.
Anything architectural like this is super interesting. We keep getting more powerful models sure. We're getting agents ok. But this coordination of many agents. This is the part that's really going to scale very fast and across many many machines in parallel. Very cool.
I wonder what would happen if the agents did not have the spec. Like if you just said "build a simple SQL database that works as a single file", what would the agent swarm come up with.
Tinfoil hat: what if the prices are artificially inflated deliberately to price out casual users being able to field large swarms of frontier model agents because it could be just too dangerous.
Love to see these crazy kinds of experiments going on. Even if this doesn't 100% work or is prohibitively expensive for now, these are glimpses into the future in the same way people were talking about coding agents in 2023 when we just had tab complete.
It shouldn't be prohibitively expensive because the more you break the work into smaller parts, theoretically the smaller a model can be that completes it.
Therefore your agent swarms don't have to be cutting edge API calls, they can be hosted on clusters of tiny cheap computers with small CPU GPU. For orgs you could basically invest in a local server that has 100s of small computers.
You can use engineers and frontier models to plan and do task subdivision. And then shit that out to the local cluster that's basically handling each small task without knowing what it's supposed to do.
The real issue is the middle. Testing, verifying interconnectivity of mid level abstractions, building that amount of tech debt at such a high rate and actually being aware in any way what it is that you've built.
I think for most MBAs, considering the US and a significant amount of global markets only care about short term market upside, and AI is still in the middle of the hype cycle, that's not a serious issue for anyone who is profit oriented.
It's only a real issue for the losers like you and me who care about sustainability and infrastructure that survives longer than one market hurricane season.
This so much!
For my current setup, the most efficient way is to use larger models for coordination, but a heretic'ed qwen 30B model for the implementations. If you build your agentic environment around specifications and test coverage tied to symbols, you can do a lot of parallelization of agent work.
If you then separate the filesystem and tool read/write access by agent roles and policies (e.g. coder not allowed to modify unit tests, tester not allowed to modify code files) then you can force them to use a centralized per symbol/per contract specification tool.
And then you can just let agents discuss issues, where the messaging threads are also tied to the same symbols.
[1] Exocomp, highly experimental: https://github.com/cookiengineer/exocomp
The twin trends of AI slop being churned out at tremendous rate and deployed in everything, plus frontier AIs becoming really good at exploiting many small security holes to pwn boxen, is going to collide in an amazing cyberstorm that was predicted by every cyberpunk novel from the 80s X-D and we are gonna have front row seats 8-D
I am afraid the folks are still busy playing with familiar toys. "Let's extract more code out of LLM's inner layers!" SQLite we already have. That's where their "spec" came from.
The interesting frontier is elsewhere: how to fit it into the pipeline to build new useful things. I felt motivated enough to write a follow-up/reply post http://replicated.live/blog/follow-up
It's cool stuff, but I think mostly in an aspirational way right now. The tired cliche hit on the tired cliche claim will be "but sqlite is in distribution" which will miss the point, that objection dissolves at the precise point where it would have operationalized: novel software of value. Is SQLite in the train set, yeah. Does that matter? Only if someone is willing to pay for Rust SQLite as something other than marketing.
It's cool but I'd rather know what their most successful users are doing, and how they measure that.
>The browser swarm from earlier this year peaked at roughly 1,000 commits per hour on Git. The new system peaks at around 1,000 commits per second.
>To facilitate this rate of activity, we built a new version control system (VCS) from scratch. Throughput was not the only reason to own this layer. Every change in the system passes through the VCS, so it is where collisions first become visible, and several of the coordination mechanisms in the next section are implemented directly inside of it.
Talk about inventing the universe to make a button.
Making a high-throughput revision control system is doable, there have been experiments here on HN with git-in-postgres etc. The question is, what is the actual bottleneck here?
I made a VCS for LLM assisted work, bootstrapping it on its own code as it was developed. In fact, if it did not really work, it would not even exist! In my experience, keeping track of things and evaluating them was the bottleneck and I focused on that https://replicated.live/blog/worktree Effectively, the question is "how to leverage human evaluation and feedback".
The Cursor people concluded their last post with "taste, judgement, and direction came from humans". Sounded surprisingly sane after "don't compile it, commit it faster" some paragraphs earlier.
People often think of tech/business as bottleneck issues ie whats my limiting factor. That's relevant to pre-ai business as humans are relatively self organizing.
The issue with AI is it has no real memory, context, long term reasoning, etc. These are a lot of what we use to determine how well we're doing something as well as why.
The "bottleneck" so to speak with ai is that it will just circularly build junk forever, expand scope infinitely, and just continue building.
At its best it will autobuild your unicorn saas. But on average it's a cancer cluster of ever expanding code with no plan.
The issue with unsupervised AI is that when you increase the throughput you decrease the amount of human supervision. So you have to have some way to deal with that.
If I'm going to give AI a gigantic task that's really complicated and it will probably mess it up 100s of times, how will I compensate for that without constantly looking over it's shoulder?
But that's mostly just vanilla models.
Some critical parts of agentic development are literally months old while most mature ones are merely a couple years old.
Critical thinking was the most critical checkpoint for models to reach.
Now its only a matter of time for the right tooling and orchestration to come along. Once agent memory and recall works well, all of those bottlenecks will go away one by one.
Back in my early career days, CPUs were clocked at 1GHz. That was 20-25 years ago.
[dead]
It's hard not read such quotes and immediately think of the Infinite Monkey Theorem.
https://en.wikipedia.org/wiki/Infinite_monkey_theorem
If at the end Shakespeare-level literature is produced, does it matter whether we arrived there by random keystrokes?
Shakespeare-level literature was produced. We are the infinite monkeies, and one was shakespeare.
How much slop you figure Shakespeare produced?
Honestly judging from the volume of his output, a ton.
It’s rare to run into prodigious artists who simply produce a few perfect pieces - as a rule those few magnum opuses are built on a hill of cutting room floor scraps.
The director James Cameron stands out because he has only produced huge box office hits. Extremely hard to do like you said, and most artists and auteurs would be happy to have one great work they can be remembered for, even if it's amidst a dozen forgettable pieces.
He's only published blockbuster hits. But I'm sure he has lots of practice films that have never seen the light of day?
Don't let your personal love of Piranha II convince you that it was a box office hit.
Sometimes the forgettable pieces come later. In JC's case, that's Avatar 2
And then there are those like J.S. Bach.
I don’t know how to feel about HN thinking one of the greatest artist of all time produced slop. You guys really think Leonardo davinci was drawing ladies with 6 fingers before the Mona Lisa?
I assure the worst of Shakespeares work is far beyond anything most writers can produce, much less the incoherent slop LLMs generate
The problem is the infinite other literature produced along the way. It doesn't work to throw more monkeys at the problem of verification, and checking for matches against existing Shakespeare plays is cheating.
don't worry, we can get another swarm of random monkeys typing and decide what ones to eliminate .
tijebs gi yo,
Right. The joke with shakespeares monkeys or borjes library is that you just do that and then filter out the books that match shakespeares works.
If you already know what you want in an infinite library, it is there. If you don't know exactly what you're looking for to the letter, you might not be able to find it.
Yes, because the folks waiting slack-jawed for the monkeys to finish will no longer have the skills or discernment to know what Shakespeare looks like.
And if in the end it does not produce masterworks, going faster and breaking things will be seen as a huge and credulous waste of time that poisoned the internet and some codebases beyond redemption.
But I’m glad someone else is running all these experiments so we can see who is right in their guess about the outcome.
Yes because the keystrokes aren't free and the world is finite. We need to pay these monkeys in bananas.
> does it matter whether we arrived there by random keystrokes?
Douglas Adams said this beautifully in the Campaign for Real Time & the poet Lallalfa.
And I just found out h2g2.com is no longer online, so no reference.
2*128 is the UUID namespace, which is astronomical. That is like 21 letters. Making one sentence of Shakespeare by random search is strictly 0% chance. SHA256 is ~43 letters then. "LLM broke all of cryptography and creates random data to fit a hash" -- are we there yet? Random search does not work at this scale. In all their cases, they played with pre-existing well-studied software, so they had an embedded fitness function. An extremely detailed one.
Monkeys have been reading Shakespeare all their lives (training). Effectively, this is LLM uncompression. Hence, "in Rust, from nothing but its documentation" -- not entirely true.
What embedded fitness fuctions are you thinking of?
Yes it will be a terrible waste of time since dedicated humans can do it faster and produce better stuff. I would prefer literature created like that instead of just random monkeys writing what they don't understand. We also lose the ability for writers and creators to get better at their craft and produce unique works instead of just randoming spamming keys for millions of years
Yes because it is akin to producing a universe of slop and trying to find the single atom of gold.
Isn't that the core fundamental of AI? Constrain the results to a latent space which has all the probable answers and then pick up whichever token makes sense based on your random sampling?
I feel like the whole concept of harness and LLMs IS to pick random tokens which make sense. Think about this as a solution space (a very complex one), then the harness and the LLM are moving towards the local minima at which the code is complete. This is akin to how we think about AI anyways.
Yes, agentic coding is machine learning. Not because of the underlying language models, but because it is gradient descent, where the test harness provides a gradient for the agent to traverse.
Seems very reasonable to me tbh, the VCS was built to facilitate human work structures, rebuilding it to facilitate autonomous work seems appropriate.
They probably just vibe coded it with ChatGPT
What would have happen if you do not commit at all?
I mean why even use VCS at that point?
imho, to be able to revert and recall as agents loose context due to new spawn or compaction.
> To test that progress, we returned to a task the old swarm had struggled with: building SQLite from scratch, in Rust, from nothing but its documentation.
Isn't SQLite's source code in its training data?
Of course it is
not in Rust. might be different enough ...
Even if no Rust code for it was seen during training, an LLM can trivially transpile SQLite's C codebase to Rust on the fly. For example, I just asked ChatGPT to write John Carmack's famous Fast Inverse Square Root algorithm in Erlang, without searching online or thinking, and it transpiled it immediately (while also extracting the knowledge in the same step). SQLite's semantics/code are stored in the middle layers of an LLM, and the last layers are able to convert it into any representation, as conditioned by the prompt. Cursor's experiment is deeply flawed because they merely extracted the model's compressed, lossy knowledge of SQLite's codebase and then just ran a bunch of tests/fixing rounds to make up for the lossiness. The claim that the agents built it from scratch is false.
Then you would expect the implementation to be structured the same as SQLite, and having glanced at the result, it looks like at least some things aren't. For example, it seems to use an operator-tree executor rather than SQLite's bytecode interpreter.
The point is that the model has the source code in its weights. Therefore it's misleading to focus on being "only given the documentation", when it could produce a SQLite implementation in rust with no documentation at all. The wording is designed to create a marketing impression that it can do spec-driven development for a novel project, but this example doesn't actually support that.
It's like a diet system advertising how much weight a spokesperson lost on their system, without disclosing that they also went on Ozempic at the same time.
The knowledge is lossy, and code generation itself is non-deterministic (temperature), so the operator-tree executor must be interference from other DB implementations, because it's uncommon to have a bytecode interpreter
To test this hypothesis you would need to look at the initial version before subsequent agents started refactoring it.
Why tho? The whole point of agentic is that it takes several passes to get to a result. If step a) is a non-idiomatic transpile, but step b) c) and ... z) make it idiomatic and performant and efficient, why look at step a)? That was just the beginning. The idea with "agents" and "swarms" and dynamic this and that is that you only look at the last step.
This strictly follows the early agentic usage, where people were complaining of code full of errors and such. But having the loop do plan-write-verify is the thing that got us from that state to today. And it works. If you look at the code and it has errors (linting, etc) then your loop is not properly setup. You should only be called upon when all that can and should pass does. Your role is "direction/taste/human-common-sense-at-a-glance".
Because the point of the test is to see what the agentic swarm approach is capable of. If your test assumes that a version of the thing you are trying to build is available in the training data, or can't determine if it matters whether the thing you're building is in the training data, then it's not a very useful test. Most of the time when you consider agentic approaches, you're not trying to rebuild something the agents involved are trained on.
I see, that makes sense. Thank you.
There's Turso which is a sqlite rewrite in Rust with some additional features. It's open source so I'd imagine it would be in the training data.
How do we know if these models weren’t trained on Turso’s rewrite of SQLite in Rust?
It seems both likely that they were and impossible to remove that code from pretraining. Doesn’t that make this just about LLM memorization of the training set? What am I missing?
> What am I missing?
They’re testing the same models on the same task but different ways of organising the swarms and the new approach works better.
This is more about automated long horizon work.
Wouldn't that be in the authors' interest to disclose, if true?
I would have loved to see more of the harness engineering shared as code. Instead we are left with only the outcome.
I guess that makes sense since the harness is the product in the case of Cursor.
I call this junk "meta-agentic engineering", it reminds me of people who have the coolest nvim configs, spend hundreds of hours customizing it but ultimately get less work done than the guy with minimal workflows, if any at all.
I look on twitter and it's just people building tools for agents to use agents, some weird customization loop going on in the LLM space right now. Ultimately these are trends pushed on us from model providers because they 10x token consumption. Its literally just BS trends to increase revenue at these companies, most of it is largely useless.
Some of this stuff is ludicrous. I finally tried /loop last week and discovered every loop iteration passes the entire context history to the model. So pretty quickly you're running a full 1M context window, without cache, likely just to check if something is ready or needs to be done. It's miserably terrible engineering unless your entire and only goal is to burn tokens.
Wait, why isn't context cached with /loop?
Because it's typically cached for 5min (1hr is a setting you have to explicitly opt into), and very few people run loops at a cadence of < 5 mins.
https://code.claude.com/docs/en/prompt-caching#on-a-claude-s...
"On a Claude subscription, Claude Code requests the one-hour TTL automatically. Usage is included in your plan rather than billed per token, so the longer TTL costs you nothing extra and only affects how long your cache stays warm. If you’ve gone over your plan’s usage limit and Claude Code is drawing on usage credits, you are billed for that usage, so Claude Code automatically drops the TTL to five minutes."
Oh! I thought Anthropic cached for one hour (by default), am I wrong about that? Or is this an OpenAI thing?
Anthropic changed their cache duration for some reason a while back.
[dead]
I never liked the term "ricing" with window managers, but this is definitely veering into LLM ricing territory sometimes. And people are getting rich off of it.
I think many of us engineers (including myself) are suffering from AI-induced procrastination. Fixing that pesky bug in production isn't fun. Building a harness/workflow that detects and fixes pesky bugs in production on its own is way more fun (especially when your employer pays for tokens), and it feels just as important, if not more! So what if it didn't actually fix that pesky bug in production on its own...
I'd rather anyone using llms to fix pesky bugs in production gets fired and I work at a shop that cares about quality and craftsmanship. There are handmade swiss watchmakers, and then there are assembly line workers in China also making watches, which worker do you think goes home happier at night (and who's company is actually beloved), you have a choice.
I don't think there is a lot test harness, given there is a golden implementation (sqlite3), with clear interface (SQL, and sqlite3 db files). Once you slop crappy implementation, then you can create one group of agents finding gap where slop doing wrong vs golden, and another group of agents filling the gap meeting the conformance.
I assume they did similar, as long as test case(or spec) can be parallelized. Having parallel execution / testing environment is crucial if we need 1000 commits/second throughput.
I would not surprised they have 1+ millions of specs.
I would bet they won't release generated specs - actually this will be a valuable asset like https://sqlite.org/th3.html
Am I reading this right? Opus + Composer did a comparable job to Fable, at ~1/19th the price, and half the LoC?
Yea, looks like it (at least for their v2). Fable did much better in V1, but once they added their tooling around it, Opus + Composer ended up doing better (>0.8 Grade on the SQLite chart) for the final product. Granted, Fable reached a 'good' result (0.7~) in a much faster timeframe.
Point is, frontier models don't necessarily are better than a swarm of non-frontier ones for these scoped problems. IMO frotnier models excel at underspec'd or more complex problems where ideation and exploration are key (and that's where they become crazy expensive).
I'm struggling to use Fable correctly. Everything I do ends up getting flagged and reverted back to Opus. Need to do some digging.
Anthropic wants it that way so they can’t be sued… the correct answer is don’t, switch to Kimi K3 for much more usage, the same quality of model, and no hidden practices :/
Today it is SQLite, tomorrow Linux?
I'm not sure of the long-term viability of these brute force swarm approaches, but clearly the AI companies dig the marketing they get out of it. But the fact that they get so far at this stage suggests we have a very interesting 2 years coming up.
This is almost a year behind Steve Yegge's first post on beads. Gas Town and Gas City provide orchestration for the swarm. So far I haven't seen a perfect implementation but this idea isn't new.
I don't think Gas Town was actually all that much ahead of the curve - see Cursor's swarm post from January this year, mentioned at the very top of this blog post.
I also don't know that beads is a very strong example that Yegge knows what he's doing - it's famously been described as (pseudo)malware.
Everything Steve yegge has done has been trash. That's why nobody is talking about beads or gas town. It was clear even in the beginning that it was a borderline AI-psychosis-fueled trash fire.
At the same time a lot of the stuff in this Cursor post sounds like things that Gastown was doing, albeit more sober and thought through
Agreed ... I'm not talking about the implementation but rather the idea that this research is novel.
I think if you're suggesting that Steve Yegge somehow invented having agents collaborate, you're way off base. People started talking about agents collaborating very, very early on, and people have been doing experiments like this for ages. There have been libraries to build e.g. graph-based (possibly multi-agent) workflows from the early days, before even structured output was a standard thing in the APIs.
The only thing Yegge did was come up with really stupid, convoluted and anthropomorphized language to describe the process and then write unhinged articles about it
Does every peice of prose published here have to be novel? is it not enough for you to read a new. perspective?
> One reading is that it was more productive. Another is that most of those commits were busywork (thrash, contention, churn).
Kudos
Even at a much smaller scale of a few dozen parallel agents working I've found real benefit to a structured hierarchy of agents. Not just keeping the context clean and focused at the right level, but if you find that something has just gone off the rails, you can more easily rip out or even just fully delete a leg of work without damaging the original design. It's also nice when you find a design change is needed and you can communicate that to the top level and have it efficiently propagate down where it matters.
I like it so much I even recorded a video walking through the process!
https://www.youtube.com/watch?v=efUmcCiRoDU (I do walk through how I do this in our app DevSwarm here)
I initially thought that getting agents to work for longer and in large groups was the future, but I'm increasingly thinking that, at least for engineering, just one thread makes more sense. The agent pulls things into context as needed. One thing that I've been experimenting with is also letting the agent remove things from context, such as files. But just adding to the context and compacting when it's full seems like it might beat a lot of more advanced options. Because the model is good, it knows what to put in the summary; just enough for the model to be able to rebuild the context from that seed (e.g. pulling in relevant files/data into the context).
Have you used this method yourself for long workflows and with contexts approaching 1 million tokens? It doesn't work very well. LLM context is nearly half unusable.
I always let GPT 5.5/5.6 sol reach compatction. I'm only offered a 258K window, but I think once they release 1M to normal plan users, it will be usable across the whole 1M. At least with Opus I could use the whole 1M. I'd argue that the more context I use, the better performance gets. It has more info already available. I don't notice degradation.
I've found that model performance starts substantially degrading after 10%.
I see with qwen3.6-35B running in Pi that it gets stuck in loops when the 262k context gets more than 30% full.
This is super fascinating, and I loved seeing testing of where exactly you need frontier intelligence -- looks like coordination / planning, but not coding right now -- the article's a bit of a tease, as we can't play with such a harness, or their new version control system or get a workable artifact out of it.
That said, I love the work on figuring out these harness coordination jobs. While there are analogs to human management there is also this enticing feeling that, since the models are broadly deterministic, we might be able to get repeatable science-type lessons about managing them with enough testing.
The issue is that the only models that could be trusted to work autonomously cost more than a human employee
Cursor is saying we need to get improve the harness and tooling, but the models are getting better. One would have to assume that will continue and that cheaper/smaller models will continue to become even more useful.
If there's a human who could write SQlite in an hour we should be paying them at least that much.
These tests are interesting but I can't help feeling that these are benchmaxxing results. "We made our harness make sqlite in rust" sounds like a much easier probelm statement than having to recreate an application like facebook. I think the concept of having to integrate is the real challenge. Re-writing code in new languages based on decent docs is a difficult but less useful measure of how AI can help replace engineers.
Imagine this: you need to build a new CRUD app but your AI agent spends 20M toks on just writing it's own custom login function, that's a waste of time and money. I would not like to have a new system developed for me in such cases. I think we need a test where the agents are more specifically tasked to find technologies worth integrating and hence write the fewest lines of code.
The nature of ML is to reach for easily measured proxy problems which can then be optimized. Why worry about a backoffice integrated micro-service you have no data on when you aren't sure that you can design and implement a rewrite of an OSS project?
This post describes ideas found in existing agent orchestrator systems since they've existed. You'll find the same "top level agent that breaks a problem down into spec / plan / implement / verify" pattern in get-shit-done or superpowers.
Is the original part the VCS that works at 1000 commits/second? Or spending a bunch of money comparing models?
Wow, people have been working hard for a long time in Turso: https://turso.tech/blog/introducing-limbo-a-complete-rewrite... to do the same cursor did in justo no time?
If we take coding off the table i don't see the case for large agent swarms. More is not always better in my experience experimenting with agents
If you can't see cases for it you may need some glasses. Replace the graph in the docs of Top-level planner -> mid-level planners -> worker leaf nodes and you get an org chart of CEO -> executives, N-levels of management, and ICs at the leafs. It directly maps to almost any office-based company that can operate completely digitally.
It's obviously in its infancy but fully agentic companies are becoming a possibility, given the right harness, models and tools
Lets separate "workers" from "decision seats", you might need a lot of workers which typically don't actual require inference and might actually just be cheap solvers but actually less decision seats. So i guess it depends on what you are calling agents in your scenario.
I find these blog posts (and the originals, with Anthropic's C compiler and Cursor's browser) somewhat funny, as if they have this enormous power to build ... but they can't build something unique or new. Like the software sucks, but look how powerful the process is (the models are indeed powerful).
And it's a bit of a shame: by virtue of their position (their embedding in the fabric of venture capitalism), it seems like they can only make a subset of things -- what they can make is dictated enormously by capital, as they are engines of capital.
Not sure the point I'm trying to make, I just find it amusing.
Perhaps the point is that it might be more worthwhile to give independent creators a billion dollars to play around with agent swarms if we want to keep diversity in the evolutionary algorithm that is the software industry high.
Because it's mostly a load of crap.
Most of these reusable, automated workflows sound really cool in theory, but in practice either they require a lot of customization (not really reusable) or a lot of babysitting (not really automated).
AI has improved my productivity significantly, I'm not a hater, but we also have to be honest about its (current) limitations. The whole "software factory" thing is just a pipe dream people sell to execs / investors / online courses.
Yes and this can be proven because what Anthropic says and does is vastly different
They say SaaS is dead and you can vibe code any SaaS using Opus and yet despite having Mythos with them they have to use JS terminal to run Claude Code.
Why don't they just vibe code CC? Why doesn't Dario have a 100000 agent swarm to run entirety of Anthropic or at least software department?
They've said for some time that Claude Code is 90% LLM generated code.
Yeah and it's still a cli tool running on JS
So it's not a brag
They could just write in Rust directly can't they? If they do that then they're doing what they preach
My app is written in TypeScript. With agents I could rewrite it in Rust, I just don’t need to so I don’t. It doesn’t warrant that level of performance and the business logic is more digestible in its current form. Given the amount of reviewing I need to do now, readability is worth more than performance. There are always tradeoffs.
Are you implying that them attempting a Rust rewrite would fail, and this is why they’re not doing it?
Anthropic was saying that you don't need to review the code anymore.
If that was the case, it just makes no sense at all to use JS. Take the free performance gain and don't read the code.
Yep they can just prompt Fable or Mythos!
Here is the prompt they can use:
Claude, migrate Claude Code into Rust. Make no mistakes!
Genuinely Fable would do a decent job.
Yep and they still don't do it! So we should do as they do, not as they say
I do somewhat agree. I can only speculate that they're pumping out new features incredibly fast and they don't want to fall a month behind the competition.
Why will they fall behind competition when any feature is just an English prompt away?
Unless they're lying that is!
True.
> business logic is more digestible in its current form.
Well, Anthropic spent many $ teling everyone that code is dead so this doesn't matter to them does it? Their business logic can be written in English!
What do you mean by “they spent $ telling everyone that”? If anything, they told everyone that to receive $ in investments. Those statements are pure PR hype.
claude code VSS - 70Gb
codex VSS - 600Mb
Also famously terrible. If not for the walled garden and the fact that anthropic can take away our ability to use custom harnesses any day, I would not be using CC. Given some of the features they add I've also started to suspect that their models are making product decisions. Not necessarily intentionally, but more in the sense that I think their engineers just let their models build and add whatever it feels like is a good idea to the codebase and then they just publish that.
Yes, that's true, but in the post, they even mention that building the spec is the scarce resource.
We're rediscovering the "architect" job role a decade or so after switching to agile processes and staff engineers largely replaced it. God help us when they rediscover UML for agents.
I feel like the scarcity of (idealized) Product Managers is even more acute.
Easily the biggest delta between a good one and a bad one I've ever encountered. It's so easy to test too. Just ask a product manager if they can give a talk on <some product they own> in 2 hours. If they respond by slacking all of the engineers that work under them to stop what they're doing and give them the info to give that talk, fire them. If they are ready to talk right now then give them a raise.
I mean, how long are these pages? Assuming 40 or so lines per page, that would be 33400 lines of spec turned into 200,000 lines of Rust, or about 6 lines of code per line of spec. That's a very detailed spec, much more than any software specification I have ever written in my career. It's not obvious to me that writing this spec would meaningfully save work over incrementally building individual features, especially when you think about what it would take for human beings to meaningfully review whether those 835 pages contain the right intent.
not only that, but what all of this misses is the fact that the specification that they used was a product of writing the software, not the other way round. It would be much, much more difficult to first write a correct specification than to first write a correct program, because programs are formal systems that can be executed to observe their correctness. You can’t do that with a specification.
Or to validate the accuracy and efficacy of the actual code produced.
but that's what we use the agent "swarm" to do .. it's also not scarce. what's scarce is imagination.
Well, the tool works well. I am grateful that Nature has seen fit that the same kind of person who can make it doesn’t also have the ability to identify what needs exist in the market because then all software would flow out of OpenAI and Anthropic and the rest of us would run about their legs and peep about. Whether product development will survive the next generation of models is still unknown.
bro it can't even make a button that doesn't look like it was designed in 2018, everything is ass. Product development isn't going anywhere.
it feels kind of like a concept car a large manufacturer would show off at a motor show. Everyone knows there is no intention to actually ship that thing, but they're exploring concepts and technology and trying to make inroads in what comes next.
The bottleneck has always been on the product side for 99% of cases. For the 1% where it's engineering, LLMs don't help as much.
Anything architectural like this is super interesting. We keep getting more powerful models sure. We're getting agents ok. But this coordination of many agents. This is the part that's really going to scale very fast and across many many machines in parallel. Very cool.
Seems an extreme system looking for a use case, but it's great to see research being done in this area
I think experiments like these are exciting, but they only work when you have the full suite against which you calibrate.
What I'd love to see is a small group of humans steering a million agents.
They have invented software time travel: Write software => then write manual => then feed manual to cursor to write the software.
I wonder what would happen if the agents did not have the spec. Like if you just said "build a simple SQL database that works as a single file", what would the agent swarm come up with.
Amazing test. I wish I could spend $2000 of compute on having my own swarm build some weird piece of software with 1000 commits per second.
Perhaps you can get hired by a company to do that. You can't compare yourself as an individual with a company with a large budget.
Can a swarm of agents write the 835 page manual?
Good news, the new manual will be 83500 pages.
That's 100x productivity right there.
I feel bad for the engineers at Cursor who have to use Grok in these sorts of experiments.
They received $60 billion to use grok and make it better. I think they going to be okay.
Tinfoil hat: what if the prices are artificially inflated deliberately to price out casual users being able to field large swarms of frontier model agents because it could be just too dangerous.
The fact that the swarm needs high quality docs to work from will keep most of the tech industry safe.
1000 commits of slop per second.
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[dead]
[dead]
[dead]