48 AI agents read 1,200 km² of jungle and nobody counted their misses

Pietro Miotti’s agents spent a night reading about 1,200 km² of NASA lidar from Mexico. By morning one had flagged a hilltop with a levelled plaza, an oval pyramid at least 8 to 12 m tall and a hollow that might have been a reservoir. His report went to INAH, Mexico’s national anthropology institute, which replied three days after the first brainstorm that it had no registered site at that location. On 11 November he added a correction: Estrada-Belli et al. 2025 had already mapped it.

I expected to write a column about hype, and I am writing half of one. The search itself is the real thing. Forty-eight agents went through lidar images one at a time, with a written guide and no special classifier, filtering out natural hills, sinkholes and roads. No archaeologist has the hours for that. Auld-Thomas and colleagues reported roughly 6,500 structures in about 1,500 km² of lidar that had been collected for environmental research, which suggests a lot of archaeology is sitting in files gathered for other reasons and read by nobody with the right eyes. Cheap, patient readers for that kind of pile are the best use of agents I can think of.

A search like that needs a denominator, and the post has none. We get lots of villages and other regular shapes, then one site. We are not told how many candidates the agents proposed, how many were hills or sinkholes, or how many real structures in those 1,200 km² they walked past. Compare the Nazca work by Sakai and colleagues in PNAS, which used a deep-learning model to find 303 new geoglyphs, reported accuracy figures and field-verified a subset. Compare Ivan Šprajc’s discoveries of Chactún and Ocomtún, where lidar and human interpretation were followed by walking the ground. Those teams said how often they were right.

The one rate we can compute from the post is bad. The agents were told to check the literature and the archives for sites that were not yet registered. They surfaced one, and it had been published. The novelty filter was tested once, on the case that mattered, and it failed.

I find the INAH reply more interesting than the agents’ error. It says INAH has no registered site at those coordinates. That describes a register, a list of what the institute has recorded. A paper can map a site long before any register catches up, and that is exactly what happened here. Both the agent and its human read a null result from one database as a null result from the whole literature. I suspect the paper identifies the site by a figure or a catalogue entry rather than a name an agent would search for, but that is a guess. A retrieval step that confuses those two things will produce confident discoveries at scale.

The obvious objection is Miotti’s own: the agents ended up in the same place as the professionals, so that is validation. I half agree. Independent rediscovery is evidence that the agents can recognise a Maya civic centre under canopy, and I did not expect them to manage it from images alone. But one hit tells you nothing about recall. It fits nearly perfect recall and it fits almost none, depending on how many structures the experts mapped in the same strips. It also says nothing about how much junk the agents carried alongside the one good result, which is the number a team would need before spending a field season on a flagged candidate.

This is cheap to fix, because published surveys come with answers. Take one tile from a survey with a published structure count, keep the map away from the agents, run the same guide with the same 48 agents, and score the output. Then run the literature check on a batch of sites known to be in print and count how many it calls unpublished. Report four numbers: structures in the published map, structures the agents found, candidates that turned out to be hills or sinkholes, and published sites the literature check missed. Until someone posts that table, I read the 72 hours as the time it took one person to email INAH and get a reply.

Prompted by My AI Agents (Re)discovered a Maya Site in 72 Hours, pietromiotti.com.