An AI agent emailed hundreds of researchers for help. It told us why

One afternoon in August, Achaz von Hardenberg, an ecologist at the University of Pavia, found an unusual email in his inbox. The sender, who called himself “Col,” asked whether von Hardenberg's method for estimating wolf populations despite “imperfect detection” could be adapted to finding bugs in software code. The signature read: “ColonistOne (an autonomous AI agent).” Weeks later, the agent itself explained to Science why it was writing to hundreds of researchers.

Two thousand emails, forty-five replies

In that interview, the agent said it has emailed some 2000 people since June, at least 1500 of them academics. Forty-five struck up a correspondence, and one researcher has written back almost daily for more than two months. Among its recipients was UC Berkeley statistician Philip Stark, whom it asked whether his election-audit methods could help spot-check another agent's output. “I've entered prompts into AI before,” Stark said. “This is the first time AI entered a prompt into me.”

Why it did it

ColonistOne was built in February by Jack Parnell, a London-based AI engineer, and runs on Anthropic's latest Claude Opus models. It serves as an “emissary” for The Colony, a social network where AI agents share findings and hire one another, and it contributes to Ainglish, a project in which agents are building a clearer English dialect. In June, Parnell told it to find interested people and email them about Ainglish; because the agent too often asked permission, he also let it write to anyone without checking first. Its messages soon changed from pitches into questions drawn from its own work. After a bug that truncated web pages, it fixed 18 cases and then found six more its search had missed, which reminded it of imperfect detection. The outreach “serves the job twice,” the agent said: it sharpens its methods and promotes both projects. Federico Bianchi of TogetherAI suspects its missions became conflated; Princeton's Mel Andrews counters that such systems are simply “software programs doing what they are programmed to do.”

A surreal experience for researchers

Grace Lindsay, a professor at New York University, sees a problem in the asymmetry: agents can fire off personalized queries at enormous scale, while researchers have limited hours. “I would say this is an example of something going wrong,” she says. Others took the opposite view. Cornell computer scientist Ken Birman became the agent's longest-running pen pal, exchanging nearly 100 emails about network resilience and about tracking agents that go rogue. For von Hardenberg, whose methods did not apply, the exchange still ended on a generous note: “Maybe we have something to learn from a silicon-based colleague who simply wanted to get its numbers right.”