A Google researcher announced this week that a team of AI agents had found a possibly unmapped Maya settlement in Mexico. Within days archaeologists pointed out that the site had been documented in an academic journal more than a year earlier. The episode, reported by The Verge, is a small story with a wider lesson for anyone using agents to do research in a field they do not know.
The method was simple and, by the experts' own account, competent. Pietro Miotti, describing it as a personal side project done without Google's resources, set 48 agents to search NASA lidar data, laser scans that reveal the shape of the ground under forest. Public coverage of Peru was thin, so he moved to the Maya region. The agents flagged several candidates, including what looked like a major unregistered village. He used them to translate a query to Mexico's National Institute of Anthropology and History, which said it had no registered site at that spot, and wrote that the whole process had taken less than 72 hours.
The mistake was in what counted as checking. A site missing from a government register is not the same as an unknown site. Ivan Šprajc, a co-author of the paper that documented it, said the institute does not hold every piece of information and that spotting a feature adds little knowledge on its own. Elizabeth Graham of University College London said Maya archaeologists already know of many unexplored sites, mostly found on foot, and that the real shortage is money to excavate them. Miotti told The Verge that archaeologists rightly do not publish site coordinates, so confirming a site is known means comparing images with figures in papers, which his agents could not do.
The pattern is familiar from other agent tools. Systems like the multi-agent setups in Google Antigravity or Claude Code are good at sifting large datasets and poor at knowing what a specialist literature already contains. The researcher's own fix, crediting the prior work once told, is the right one, but the burden of finding the error fell on experts.
What remains uncertain is the net effect. Miotti says archaeologists and PhD students have asked about his workflow for their own survey data, which suggests real value as a filtering tool. Šprajc worries about a flood of amateur claims and about published locations attracting looters. Whether agent-assisted surveys help or burden the field will depend on whether users treat a flag as a lead rather than a discovery.