48 AI Agents Combed Through NASA Lidar Data and Found an Unknown Maya City

Usually, "AI found" in a headline means a slide deck with pretty charts. This story is more down-to-earth, and therefore more interesting: one researcher, public data and a whole swarm of agents instead of a single big model.

What happened

Researcher Pietro Miotti wrote on his website that, using 48 AI agents, he discovered a previously unknown city of the Maya civilization in Mexico. The source material was publicly available NASA data obtained by laser scanning of the Earth's surface, that is, lidar data.

In this data, the algorithms recognized three signs of an ancient settlement: a leveled plaza, an oval pyramid and a water reservoir. Mexico's National Institute of Anthropology and History confirmed that its database listed no registered archaeological site at the given coordinates.

How it works, with a simple example

Lidar sends out laser pulses, and some of them pass through vegetation and bounce off the ground. The result is a terrain map in which the forest has seemingly been removed. Ancient platforms, mounds and bodies of water show up on it as shapes that are too regular, in places where nature does not draw geometry.

Imagine you have to check thousands of square kilometers of such maps. A human gets tired and misses half. Agents do not get tired: each looks at its own area or its own type of anomaly, and the results are then merged. How exactly Miotti distributed roles among the 48 agents is not disclosed in the source, so we will not speculate about the architecture. But the idea itself is familiar to anyone who has built a data-labeling pipeline: many cheap passes instead of one expensive one.

Where the catch is

The main caveat: for now this is a candidate, not an excavated city. The institute's confirmation says only that the registry is empty at these coordinates. That does not mean nobody knew about the place: local residents or researchers who never entered the site into the database may have known about it. The report mentions no fieldwork, and the word "city" should be read as a working hypothesis.

The second problem is false positives. An oval shape on a lidar map can turn out to be natural terrain, an old farming terrace or a processing artifact. If the agents were asked to "find a settlement," they have an incentive to find one. The source has no data on how many candidates the agents discarded and how many they flagged in vain, and that figure is exactly what would show how reliable the method is.

The third caveat: we know about this from the author's own account and the institute's comment on the registry. The material contains no independent publication describing the methodology.

What practitioners can take from this

First, the value here is not in model magic but in the combination of "open data + cheap parallel search + an expert for final verification." This scheme can be carried over to other tasks: anomaly detection in satellite imagery, infrastructure audits, archive labeling. If you have a large body of public data that nobody has reviewed in full, it is a candidate for this kind of processing.

Second, build in external verification from the start. In this story, its role was played by the national institute, which checked the coordinates against its database. In your product, such a checker could be a registry, an expert or a field test. Without it, an agent will produce convincing but unverified findings.

Third, measure and publish the false-positive rate. For a customer, what matters is not "we found 1 object" but "of N candidates, M were confirmed."

Bottom line: this is not a scientific sensation but a good example of how the agent approach works on real public data. Let's wait for field verification; then it will become clear whether the agents found a city or just a pretty hill.