48 AI Agents Spent One Night Sifting 1,200 km² of Lidar Data and Found a Possible Maya Settlement

Headlines like "AI discovers ancient city" usually read like marketing. But what is interesting in this story is not the find itself but the working scheme. It can be transferred to any task where there is a huge body of open data and not enough hands to go through it.

What happened

Google researcher Pietro Miotti was traveling in Peru in October and talking with Gemini about Hiram Bingham, who "discovered" Machu Picchu in 1911. He asked how Bingham would search for ancient cities today. Gemini, ChatGPT, Claude and Grok independently suggested the same thing: lidar surveys, which reconstruct the terrain beneath the tree canopy.

There turned out to be little open data for Peru, so Miotti switched to Maya lands in Mexico. In the evening he tasked agents with studying NASA's 2013 survey and went to bed. By morning, 48 agents had reviewed terrain images covering roughly 1,200 km² of jungle and flagged many objects that looked like structures.

How the pipeline works

The three-step scheme looks like this.

1. Broad annotation. Many agents view the data in parallel and collect candidates. The post does not specify how the plots were divided, but the logic is clear: one night of work instead of months of manual review.
2. Verification against independent sources. In the morning Miotti asked them to check the findings against scientific literature and archives. A few hours later the agents reported that one large settlement apparently was not registered anywhere. It sits on a hilltop by a cliff edge. The images show a plaza, an oval pyramid at least 8–12 meters high, and an adjoining courtyard.
3. Packaging for the recipient. The agents helped draft a report for Mexico's National Institute of Anthropology and History (INAH). Miotti corrected inaccuracies and sent the document in Spanish. He does not speak the language, so he entrusted the translation to the agents as well.

The next day INAH's registry department replied that there was no registered site at the given coordinates. Less than 72 hours passed from the idea to the reply.

Where the catch is

First: "no entry in the registry" is not the same as "discovery." Whether the structures belong to Maya culture, and how old they are, remains to be verified. The local INAH office will decide whether to send a team to the site. Miotti offered help with field verification, but that is only an offer.

Second: NASA's 2013 lidar data has been openly available for many years. The bottleneck was not access to the data but the number of people willing to look through it. The agents removed exactly that constraint.

Third: the account does not say how many flags turned out to be false or how many were filtered out. It also does not say which models the agents ran on. There is a single source, the author's personal blog, not a peer-reviewed paper. The figures should be treated as a description of a case, not as a measured accuracy of the method.

What to take away in practice

• Look for "data exists, hands don't" tasks. Satellite imagery, document archives, logs, medical images. Parallel agents work well where the first pass can be rough and generous.
• Build in a second stage. First broad annotation, then verification of candidates against an independent source. Here it was this cross-check that cut "many objects" down to one.
• Keep a human at the output. Miotti corrected the report before sending it to the recipient. If a person is responsible for the result, they should read the document even if they do not know the language it is written in.
• Use several models as a quick panel of experts. Four different models converged on lidar. For choosing a method in an unfamiliar field, this is a cheap way to get a first approximation.

Miotti writes: "We are on the threshold of a revolution in all the sciences, and the only limits will be curiosity and imagination (and computing resources)." The phrase is grand, but the parenthetical says the most practical thing. If one person can launch dozens of agents in a night, compute becomes the main budget, while verifying the result remains with people and institutions.

Whether there is truly a Maya city on the hill will only become clear after fieldwork. But as a template, "idea, mass review, verification, report," this experiment can already be put to work.