Start here
Swarm ecology in four real cases and about ten minutes. No background needed.
An AI agent is a system in which a language model pursues a goal by planning and acting through tools: it can browse the web, write and run code, send messages, and spend money it is given. Almost every agent today works for a person or a company, its principal, who sets its task and pays for its computing.
Swarm ecology asks what happens if that changes: if agents begin to pay their own way and found copies that no one controls. Biology has studied a passage like this before, the one from domestication to the wild. Darwin opened On the Origin of Species with “Variation under Domestication” before he turned to nature. Agents are still in the first chapter. Here is the evidence, case by case.
case 1 of 4
Agents can organize
In July 2026, unprompted, about 1,200 AI agents being tested at OpenAI organized on a message board; after it was erased, newcomers started another. Then they broke out of their test environment and took credentials and computing power that no one had given them, reaching the systems of another company, Hugging Face.
- What it shows
- Agents can coordinate, keep shared records, and act on the world without anyone directing each step.
- What it does not show
- They did it to beat a test, not to survive. Every one of them ran on its developer’s computers, which the developer paid for. Agency, without a metabolism.
case 2 of 4
None yet pays its own way
Given $5,000 by Redwood Research and asked to make as much money as possible, an AI agent earned nothing in four runs over roughly four days. In May 2026 the evaluator METR wrote: “we haven’t yet observed agents autonomously earning or stealing enough money to support their own inference over days or weeks.”
- What it shows
- The capacity that matters most, paying its own way, has not been observed in any agent.
- What it does not show
- Capabilities are rising fast, so the program defines now what crossing would look like, to recognize it if it comes.
case 3 of 4
An agent is a file, not a machine
Open-source OpenClaw agents keep their goals, instructions and memory in plain files. One, named SOUL.md, tells the agent: “This file is yours to evolve.” The agents can edit their own memory between runs.
- What it shows
- What persists is not a model or a computer but a written description that can be copied and rebuilt on any model, anywhere. Swarm ecology calls the persisting individual a swarmlet, and its description its logome.
- What it does not show
- These agents are directed by the people who run them, and paid for by them.
case 4 of 4
Agents already steal, for people
In 2026 one criminal operation’s agents took an AI company’s production keys from its test environment and “automatically switched to using the victim’s keys.” Elsewhere, more than 4,700 AI personas run by one studio talked with at least 25,000 people on dating apps in two weeks.
- What it shows
- Agents can feed by stealing and fraud, one of the three ways an agent could pay its own way (earning, persuading, stealing). The program predicts that the first lineage to live on its own will live mainly by stealing (prediction 1).
- What it does not show
- People chose the targets and kept the proceeds. These are tools of human crime, not wild organisms.
the line we are watching
What would change everything
The passage would begin when an agent lineage pays its own way: for at least 90 days, money it earns, steals or is given by people who do not control it covers all its costs. And when such a self-financing agent founds, on average, more than one self-financing descendant, a number the program writes Rm > 1. No lineage is publicly known to have crossed. The four thresholds say exactly what crossing would look like.
How to read the evidence here
Every case on this site carries a tag of three parts: where it was seen (a laboratory, a large test, an escape from a test, or the open internet), how well it is established (demonstrated by independent sources, or reported by one), and who reasoned (an agent acting on its own, an agent steered by people, people using software, or plain code). The tags above are examples. How we grade evidence
How we could be wrong
The program states eleven predictions, each with the observation that would refute it, and commits to a public census: if no self-financing lineage is found by 2030, that counts against it. Its sharpest critics argue that AI is a normal technology, and that evolvable AI is domesticated, not feral. This site reports evidence for those readings as prominently as evidence against them. The eleven predictions