48 AI agents flagged a Maya site hidden in jungle lidar

Looking for ancient ruins in jungle lidar normally takes a human team months. Pietro Miotti, who says he is not an archaeologist, did it over three days in October. He ran 48 AI agents at 1,200 km² of public lidar data. His write-up reached 911 points on Hacker News on 11 November. On the same day he updated the post. A Mexican team had already mapped the site in a 2025 paper.

The target was one Maya civic centre hidden in a Classic-period landscape. Miotti turned the lidar strips into images. He then asked 48 agents to scan them one by one. The agents looked for pyramids, platforms and plazas. They rejected natural hills, sinkholes and modern roads. He checked their best candidate against point-cloud profiles. Those profiles show the ground under 10 to 17 metres of forest canopy.

Three days from an idea to an official reply

Miotti writes that the whole run took less than 72 hours. The agents flagged a hilltop with a levelled plaza. They also flagged an oval pyramid 8 to 12 metres tall. Beside it were a raised courtyard and a flat-bottomed hollow that could have held water. Miotti wrote a short report and translated it into Spanish. He sent it to Mexico's National Institute of Anthropology and History on 20 October. INAH replied on 21 October: no registered site at that location.

That reply looked like a discovery for most of the day. The 11 November update reverses it. Miotti writes that the site "had already been mapped in Estrada-Belli et al. 2025". He had not reached that dataset before sending the report. The site is real, the ruins are there, and the archive that INAH checks did not know about them yet.

What a reader of this site can take from it

The lesson is in the pipeline, not the find. One person used four models to brainstorm the approach: Gemini, ChatGPT, Claude and Grok. He then ran a fleet of agents in parallel over a large area. The work the agents did was patient and easy to check. Read an image. Flag a pattern. Pass the hit to a second step that holds the raw point cloud. A single agent would have taken too long. Forty-eight of them answered within a day.

The update also shows the second half of the pipeline. A check against the current published literature was the step that would have caught the old mapping. Miotti ran it after the fact. A team building a tool like this would run it first. The report would then say what is new and what is already in the record.

The score on Hacker News is the signal. 911 points on 11 November means the audience for the method is wide. Readers wanted the write-up, not the title.

Source

How my AI agents (re)discovered a Maya site in 72 hours, by Pietro Miotti. Hacker News discussion.