48 AI agents analyzed 1,200 km² of lidar overnight and found an unregistered Maya settlement

Agentic AI usually comes up in discussions of coding, mathematics, and physics. This story is different: someone who freely admits “I’m not an archaeologist” went from a passing thought while traveling to an official response from the Mexican government in three days. Let’s look beyond the romance of discovery and examine how the workflow was put together, because it can be reused.

How the pipeline works

It all began in Peru. The author was asking Gemini about the history of Machu Picchu and recalled that Hiram Bingham found it in 1911 almost by chance, when a local farmer led him to the ruins. The question was simple: what would Bingham have done with today’s tools? Gemini, ChatGPT, Claude, and Grok all suggested lidar. In satellite imagery, ruins beneath the jungle look like ordinary trees, while lidar sends laser pulses through gaps in the canopy to map the ground beneath.

Peru proved a dead end. Open lidar data there is scarce, while commissioning a survey from a specialist company means permits, money, and bureaucracy. So the author switched to the Maya, whose territory has much better coverage in publicly available data.

The next steps were straightforward. Survey strips covering roughly 1,200 km² of jungle were converted into ordinary images. Then 48 agents worked through them following detailed instructions: look for pyramids, platforms, and plazas, and discard natural hills, sinkholes, and roads. No specialized classifier was trained; the agents literally examined pictures. Imagine giving an annotator a detailed checklist and a thousand maps, except there are 48 annotators, and they work while you sleep.

Searching, checking for novelty, and verifying the raw data

The most useful aspect here is that the work was split into separate passes.

The first pass cast a wide net for candidates: in the morning, the author found numerous villages and regularly shaped features in the reports. The second pass asked a different question: “Is this already known?” The agents searched academic literature and archives, narrowing the entire list to one large, undocumented settlement. The image shows a leveled plaza, an oval pyramid at least 8–12 meters tall, a raised courtyard, and a flat-bottomed depression that could have been a reservoir.

The third pass returned to the original point cloud. The agents checked that the shapes belonged to the ground rather than being artifacts produced by trees: beneath a canopy 10–17 meters high, the surface traces the same plaza, pyramid, and courtyard.

Finally, an agent suggested the next step: write to INAH, Mexico’s National Institute of Anthropology and History, and explained the procedure. The author revised the report by hand, and the agents translated it into Spanish. On October 21, three days after the initial idea, the institute replied that it had found no registered archaeological sites at the specified location.

What remains unproven

Skepticism is warranted, and the author acknowledges it.

• There is no confirmation. INAH’s response means “nothing is in the registry,” not “nobody ever knew about it.” Local residents may have known about the site, and some data may never have made it into publications.
• There are no performance figures. We do not know how many false positives the agents produced, how many candidates were eliminated, or how much the run cost: the author mentions only “modest computing resources.”
• The interpretation is uncertain. “City center,” temple, and aguada are hypotheses reached after discussions with the models, and the author concedes that the depression could simply be a pit.
• This is a personal account with no peer review. The decision on a field expedition rests with the local INAH office, because any work at archaeological sites in Mexico requires authorization.

What you can apply yourself

The most transferable lesson concerns the workflow. If you have a large collection of visual data—satellite imagery, scans, maps, or nondestructive testing data—first check whether you can render it as images and give it to multimodal agents with clear instructions. That is cheaper than assembling a dataset and training your own model.

Separate discovery from verification. One pass should generate plenty of hypotheses; another should be methodical and exacting: is this already known? Does the raw data support it? Those passes are what turned an overnight list of features into one compelling candidate.

And remember that the agents produced a list for an expert, not a verdict. The final say and legal responsibility remain with people, and the author was right to go straight to a government institute. If an expedition takes place, it will reveal how many entries on that list are actual discoveries.