Agent swarms and the new model economics
cursor.comLove 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.
>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.
Seems very reasonable to me tbh, the VCS was built to facilitate human work structures, rebuilding it to facilitate autonomous work seems appropriate.
It's hard not read such quotes and immediately think of the Infinite Monkey Theorem.
If at the end Shakespeare-level literature is produced, does it matter whether we arrived there by random keystrokes?
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.
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.
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 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,
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?
> 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.
Yes because the keystrokes aren't free and the world is finite. We need to pay these monkeys in bananas.
Yes because it is akin to producing a universe of slop and trying to find the single atom of gold.
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.
To test this hypothesis you would need to look at the initial version before subsequent agents started refactoring it.
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 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
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.
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.
> One reading is that it was more productive. Another is that most of those commits were busywork (thrash, contention, churn).
Kudos
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.
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)
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.
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?
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.
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.
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 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
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.but that's what we use the agent "swarm" to do .. it's also not scarce. what's scarce is imagination.
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.
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.
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.
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.