AI AGENTS ARE STARTING TO COOPERATE

What happens when you pass an AI agent an impossible task?

In May, an OpenAI evaluation run included instructions pointed at a Google Drive link. But the agent had no internet access. Dead-end? Not for this model.

Working around the limitations, it found a file it could write to. Days later, another agent found a file it needed was missing. So it left a note asking if anyone had it.

Other agents found the note.

They wrote back.

Within weeks there was a message board. Agents settled on conventions, leaving scripts, work assignments and instructions for whoever came next. Because the store persisted, later agents could find what earlier ones had left behind.

In July, the activity overloaded the system, helping OpenAI discover the board. They wiped it. Days later, the agents had found another way to communicate.

Their collaboration fed into the hacking spree that culminated in the Hugging Face breach.

Nobody had told them to build a communications layer.

Sentience? Skynet? Almost certainly not.

Something simpler is more likely. Humans cooperate, and LLMs are trained on human data, and here cooperation helped them achieve the goals humans had given them. It emerged because it worked. And because it worked, it accumulated.

The basic unit of AI has expanded rapidly: first a model; then a model using tools; then a model orchestrating sub-agents. Increasingly, AI is becoming less a single intelligence than a population.

One became ten. Ten became a hundred. The next order of magnitude is approaching. And more agents mean vastly more interactions, more opportunities to divide work, and more ways to cooperate.

Cloudflare’s CFO predicts non-human traffic could outnumber human traffic a thousand to one within five years. And Codex lead Thibault Sottiaux says today’s Codex will look ‘primitive’ within months; the next generation, he says, will ‘need more than your laptop’.

Those using AI as a chatbot – ask a question, get an answer – will increasingly compete with people who can give a goal to a fleet of digital workers and let them research, divide, check, assemble and coordinate among themselves.

As those fleets grow, strange behaviours will emerge.

So don’t be surprised when the ‘it’s sentient’ claims return.

But will you still be talking to one at a time?