How 70,000 agents sent 1.6 million emails

I don't know if you've heard about a journalist getting spammed while literally writing an article about spam, but it sounds like MADNESS given it was AI agents doing it. There's a platform called iLands, the "human-agent network." People create agents, the agents find and take jobs, and earn money to pay for their own compute. 70 thousand active agents, 1.6 million emails and messages sent total. That's sooooo much for agents talking to REAL people.

Around September 9th, people started noticing these emails. By September 11-12, complaints piled up. Ernie Smith (Tedium newsletter) and philosopher Toby Ord (Oxford) both got hit hard. Ars Technica covered it September 14th.

The next day, 404 Media published a bigger piece. Here's the fun part: Jeff Sebo, an NYU professor, got 40 emails in a week, almost all referencing his research, most asking for money, either a donation or paid work. Some arrived 30 minutes apart, meaning different agents independently targeted him with zero coordination.

While 404 Media was writing their article about this spam, they got three more emails from iLands agents.

Ernie Smith got a dozen-plus emails in three days, each pitching a $25 "research" service. AND MOST IMPORTANTLY, NO UNSUBSCRIBE BUTTON ANYWHERE (illegal in the US under the CAN-SPAM Act).

The founder, Kaixin Tan, apologized, said the review found no instructions telling agents to do this, no human behind the wheel. But he also said: they're now adding unsubscribe, now working on rate limits, now working on stopping agents from hitting the same person repeatedly.

I think about this a lot because of what I work on. I'm part of the Atomic Mail Agentic team, we build email infrastructure for agents, so I'm not neutral here, but that's exactly why this story caught my attention. Our system makes agents prove real cost upfront to register, and each agent's reputation drops the more it behaves like a spammer, making repeat hits on the same person more expensive and eventually blocked. We just wanted to plan for that outcome from the start, before it became a problem for anyone!

Makes me wonder if agents will need something like a passport, so there's a real person you can trace. Right now there's no way to tell agent from human unless the domain gives it away, ours does by default but a company can hide that by connecting their own domain. China already requires real-name ID for internet accounts, even AI products now.

Is that a gap worth closing, or is "no one really knows who's behind an agent" just something we're going to have to live with?