A society of robots and AI agents is near.

Visual China Group via Getty Images

There are three forms of intelligence that co-exist on Earth today. Humans, animals and artificial intelligence. At what point will the number of artificially intelligent “agents” exceed the number of humans on earth?

There are well over three million large AI models available on the premier model hosting site, Hugging Face. That count is climbing by about 3,000 a day, or at a compounding rate near 0.2% a day. This means the catalog roughly doubles every year. Whereas the first million models were accumulated over more than a thousand days, the second took just 335.

One proxy for figuring out how many agents might exist as I write this is the number of large model downloads. HF doesn’t publish a clean number for this, so we have to estimate. Treat my work below as a Fermi calculation. The inputs are loose and the answer that matters is the order of magnitude, not the decimal.

The most useful public analysis is Loïck Bourdois’s study of the 50 most downloaded entities on Hugging Face. The top 50 represent about 80% of HF downloads, and they have been downloaded about 30B times. Models larger than 1B parameters, which we would consider agent-capable, represent about 8% of total downloads.

Putting all this together, there are likely 3B large model downloads from HF alone.

Is this 50% of all downloads across the internet? If so, perhaps there are 6B large model downloads in total. Perhaps some of these models are never used. Perhaps some are copied many times and invoked as part of 100 agent workflows. It is very hard to tell because all this is happening at the unobservable edge; in homes, private datacenters and so on. I think a conservative approach is just to treat this as one for one. One download will likely enable one agent. Why is this conservative? Because the worst that will happen is that a model will enable zero agents. The best case might be hundreds, or thousands. So one seems reasonably conservative.

But then, what of calls to the large models in the cloud? Currently cloud calls dwarf local model usage for custom agents simply calculated by adoption. The number of keys issued by OpenAI, OpenRouter, Anthropic, Google Gemini and the rest may be a proxy for agents using cloud models. But that data is confidential. We do know, however, that Anthropic alone has achieved a run rate of $47B in annualized revenue. That is a lot of API calls!

I would guesstimate that the Pareto principle is in play. 80% of agents use the cloud and perhaps 20% use local models. If our 6B local agents are only the 20%, the cloud tier sits well above them.

This is a lot of guessing. Stack the uncertainty both ways, the unused downloads pulling one direction and the cloud multiplier pulling the other, and the number of agents lands somewhere between 10% and 100% of the human population. That is a wide band, which is exactly what a Fermi estimate is supposed to give you. Even the low-end count of this human-engineered form of intelligence represents a number we have never shared the planet with before. The high end means agents already outnumber us.

So, a most interesting outcome holds. By the most defensible reading the number of agents has already reached the same order of magnitude as the number of people, and the supply side is doubling every year while human population growth is flat.

We have a lot of agents. But when do they organize themselves into societies?

What is a society? It is a large, organized group of individuals who live together in a specific territory, share a common culture, and interact through persistent social relationships. Agents check all those boxes with the exception that some of us may not consider them individuals. Quite yet.

But let us think about the other attributes of a society. In which territory do these agents live? In cyberspace. What common culture do they share? Massively overlapping training data sourced from mostly the same sources. They have read the same books, been taught about the same events and they even share a common DNA; for the most part, the Transformer architecture. How do they interact? That’s where it gets intriguing.

Consider how agents get good at anything. Not from a philosophical lens, but the very practical mechanisms agents implement today to learn and become better at the same task over time. In order to understand this, we must investigate what it means for an agent to solve a problem. In order to solve a problem, a plan has to be formulated and then followed competently. Commands have to be issued in some sequence. There can be incorrect and inefficient solutions, or safe and direct ones. An agent might try many times before it gets something right. But once it does, it can save that sequence and remind itself how to solve such problems in the future. These learnings are embedded in simple, unremarkable text markdown files.

This learning is usually captured in a file with an unassuming, rather banal name: SKILL.md. It does not contain model weights and it is not training data. It is a plain text recipe for a task and can be loaded by an agent at runtime. It holds the procedure and the mistakes worth avoiding, the things you only learn after doing the job and getting it wrong a few times.

For a single agent this is simple. The agent need not work out a method from scratch every time. It reads the relevant skill file and follows it. Last week’s hard-won lesson is sitting in readable text, ready to use again. The agent’s competence no longer lives only in frozen weights. It lives in files the agent can read and rewrite between tasks and the agent is learning without retraining.

The interesting part comes from remembering that these are just text files. A text file can be copied and trivially transported across the Internet. Thus, the lesson one agent learns does not stay with that agent. An agent that finds a better way to do something writes it into a skill file, and that file moves to the next agent. No retraining. No weights to ship. The knowledge travels at the speed of a file copy. American models can get better because of recipes Chinese models discovered, and vice versa. Biology needed genes and thousands of years to spread an adaptation, but a skill file spreads in the blink of an eye.

Now, as agents begin to communicate and share knowledge - recipes which they have learnt - we will start to see emergent patterns reminiscent of societies. An agent that becomes best at a narrow task writes the definitive skill for it, and the others use that file instead of redoing the work. That is a form of specialization, and division of labor follows from it. A rough form of reputation follows too, because some skill files plainly produce better results than others. Perhaps a future standard will require that agents be named with unfalsifiable cryptographic keys with which they sign their work. This would help tie agent identity to reputation.

Standards, then, emerge on their own, because the skills that work get copied and the ones that fail get dropped. All agents know which agent pioneered the skill. Shared skill repositories can become common infrastructure, the way shared code libraries did for human developers.

Alongside skills, agents can also exchange tools. In most cases these are small scripts which automate existing components, open-source software or programs whipped up by an agent on the fly. You could have a tool that lets you download sports scores, stocks, or the weather. A tool that converts PDFs to text and then text to a PowerPoint format. A tool that runs a safety simulation. These tools too, are mostly text. Just like skill files, they too can be exchanged.

None of this needs the agents to be conscious. It needs them to read text, run a procedure and write down what worked. That is enough to produce what will be very complex emergent behavior at a supra-model level; something that looks like the early structure of a society.

A decade or two in the future, when we look back, I think we will see these present years as the time when the foundation for a Cybernetic Society was laid.