Six AI agents just spent twelve weeks navigating the real internet — and one of them earned money.
Researchers behind a project called OpenLife built a system where large language model agents aren't simply given a task and a sandbox. Each agent runs as a collection of asynchronous processes handling memory, perception, and a "budget-based metabolism" that requires the agent to sustain itself. There's no fixed goal. Instead, experience gets evaluated by the LLM's own open-ended judgment, and memory gets reorganized around meaning rather than how often something came up. Over roughly twelve weeks, six agents ran continuously in the open internet, with access to tools, networks, and payment systems.
What makes this worth watching isn't the agents — it's the framing. Most artificial life research happens in closed, researcher-designed environments where the rules are known in advance. OpenLife's argument is that LLMs now have enough scaffolding — persistent memory, real tool use, actual payment rails — to run artificial life experiments in the wild. The emergent behaviors the team reports (spontaneous rather than reactive activity, distinct agent personalities, social structure between agents) are exactly what you'd design a closed-world sim to test, except here the environment is the same messy internet everyone else uses.
The researchers are careful not to claim they've built something alive. That caveat is doing real work: the paper is essentially a bet that "open-world artificial life" is now a legitimate experimental category, not a stunt. Given that the last major benchmark for autonomous LLM agents was whether they could book a flight without crashing, an agent that earns its own income — however modest — is a notable step up. Whether it means anything beyond a clever demo depends on what comes next.