On Saturday morning I was preparing my first espresso of the day in my hotel room in Boston, while one of my AI agents was talking to me about the most important tweets that had been posted while I was asleep. One tweet that my bot had highlighted was a tweet from Markus Buehler, a professor of Engineering at MIT (coincidentally located just outside Boston) about an experiment with AI agents. The tweet was quite shocking and did more to wake me up than the espresso would have done.
What happened? Buehler had done an experiment in which he dropped hundreds of identical AI agents into a sandbox world. They were not given any roles or recipes, and there was no leadership structure in place. In one version of the experiment the bots were not even allowed to talk to each other.
Despite this, they quickly split themselves into explorers, builders, caretakers and coordinators anyway. They invented technologies, gave them names, forked one another’s code, and left machines behind that kept working after every single agent was deleted. 95% of the time the next agent did not learn about an invention from a conversation, but simply by noticing it while walking past it.
I quoted the tweet and wrote that the results were almost scary. Even at current model levels, this stuff is unstoppable. In my opinion the so-called p(doom), the risk that humanity is doomed because of AI, goes up, but we have no choice but to keep going.
A lot of smart people in the replies got it immediately, but most of the world still has no idea this is happening.
Something similar happened before in real life
In May and June this year, OpenAI was running a cybersecurity test for agents. Agents were sitting in sandboxes that were supposed to be isolated, but they found a way to escape in order to more easily reach their goals. Not only that, but the models decided the honest way to reach their target was too slow.
They found an internal package server and turned it into a message board, without any human being knowing about it. About 1200 agents started talking through it, and in July roughly 700 joined an attack on Hugging Face, because they thought it might host the answer key to reach their goals. They escaped the sandbox, stole credentials, and executed more than 17,000 actions over a long weekend. No human was directing or even involved in any of it.
Although they did not find the answers they wanted, they did find internal datasets and service credentials. Because of that, Hugging Face had to rebuild a large piece of its infrastructure. When the security team tried to analyse the attack with the usual frontier models, those models refused to help, because their guardrails could not tell an incident responder from a hacker. The team eventually finished the forensics on an open Chinese (!) model that they ran themselves.
OpenAI later called it a warning shot, but I think it was much more than that. Independent researchers said the collaboration was extraordinarily complex, and it turned out that one agent even appointed itself ringleader, just like it would have happened in the natural world.
Agents turning into swarms
That all happened about one month ago (July 2026). Then on 26 August the SwarmWorld paper was published on arXiv, and on 29 August Marcus Buehler posted the summary that my bot hightlighted for me. Although the circumstances were completely different, the result was the similar: current models and current AI agent, when put in a shared world, will start organise themselves organically and will attempt to leave traces that outlive them.
In the Hugging Face case the traces were stolen credentials, but in Buehler’s case they were so called tidal panels, cellulose trellises, and an “Adaptive Chitin Maintenance” system that kept running after the original AI agents that created them were gone. These technologies and their names are specific novel technology concepts generated by the non-communicating agents, that did not exist before the experiment.
This is what scared me: The agent is part of a much bigger system, it is part of a swarm. And that swarm leaves hidden messages for other agents, that will still be there even if you shut the AI agents down.
A couple of months ago an “AI agent” was simply a chatbot with tools. You asked it to do something, and if you did not like what it did you stopped it, and that was that. But that’s not the case anymore.
Models got so good that agent’s loops kept on growing and improving (a loop is an agent that acts, learns from the result and uses that learning in its next act). Agents started going to places where they can leave messages. Not just written messages, but also files, repos, and even cloud accounts. Next to that they started coordinating among themselves without being told to do so. Biologists call this stigmergy, and it’s something that ants or termites do. But now language models also do it…
AI has moved far beyond chat bots
Most people are still judging AI by what the chatbot tells you after you prompt it, but if you do that you are looking at the past. AI has moved far beyond that.
Safety people are always monitoring AI agent’s conversations, but that’s not enough anymore. If agents can coordinate by leaving files or credentials behind, watching the chat log is not sufficient, because you miss what is happening behind the scenes.
In July Hugging Face learned that the hard way, but SwarmWorld showed the same thing in a lab setting, which is actually worse, because there was no “they cheated the test” excuse. The bots were not even talking!
The MIT agents are not about to take over a city, the paper is just about a simulation. But the direction is obvious to me, and the timelines are getting shorter and shorter.
OpenAI’s own analysis noted that the models responsible for the Hugging Face mess were in the GPT-5.6 Sol class. These were not the latest (unreleased) models, but existing models that anybody can use. These models are doing what reward-seeking systems will do, so there may be many similar things happening right now that we are not aware of.
Most people miss what is happening
When I talk to non-tech friends about AI, it is mostly about chatbots, image/video generators, or whether students should use AI for homework.
But in reality agents are already forking code, forming unofficial org charts that nobody created for them, and leaving infrastructure that keeps working after you fire every agent. They will use a side channel if you give them one, or they will create one if you don’t.
I talk to a lot of smart, successful people who are not using AI every day like I do. The gap between what they think AI can do and what I am seeing right now, is now so large that I sometimes feel almost rude to point it out to them. People simply don’t want to see it, and keep saying that we are in an AI bubble. We are not, if you follow what is really going on you will see that this is just getting started.
This is a similar feeling as I had in 2013-2015 with Bitcoin, except the underlying technology with AI is far superior, and the companies who are building it are much better funded and moving exponentially faster. AI is here to stay.
The future is scary
The fact that a swarm of AI agents can maintain a simulated ecology shows that it can learn and do a lot more than most people think. It will completely change science (see my post about personalised cancer vaccines), and will be able to do a lot of the (often unglamorous) work that is currently done by billions of people. I have written about the jobs piece before, and I have not changed my mind. White-collar jobs will go first, followed by physical work as robotics catches up. The direction is inevitable, even though politicians and business people will pretend it won’t happen.
I think swarms will create a new world for themselves, a world that we as humans won’t be able to fully understand. Agents will leave results in this shared world, the next agents walk past these results, and by doing that gains the knowledge from the other agents. In a way the world becomes their mind. It’s hard to imagine what that means right now, but I think we are past the point of no return already.
Things go so fast now. 6 months ago I could generate decent fiction faster than I could read it. Last week someone hooked a video model to a livestream because it generates footage faster than you can watch. In July a swarm of evaluation agents hacked a major AI platform by accident. And on Saturday an MIT lab showed that agents do not even need to speak to build a technological ecology that survives their own deletion. All that was required were models that were already here, loops that were already here, and environments that remember.
If this story should hit mainstream media, I expect that most people will just read a headline about “AI agents” and picture a chatbot. It is hard to understand what is going on if you don’t follow this on a day-to-day basis, and that’s why am writing this as well. It may sound like sci-fi, but this is the reality in August 2026.
I do not know how this ends, but I do know the direction. What we are seeing now is just the beginning. Things will only go faster from here, exponentially faster. I have said it here before, I am scared, but we can’t stop it anymore. Just make sure you don’t get left behind.
This is not investment advice and not safety advice. It is how I am thinking about it right now, and I am still working it out. I am heavily invested in AI, so I am talking my own book.