48 General-Purpose AI Agents Lock Onto an Unregistered Mayan Site in Mexico in 72 Hours

An engineer used models including GPT-6.1, Gemini, Claude, and Grok to drive 48 parallel agents that analyzed public NASA LiDAR data within 72 hours, covering about 1,200 square kilometers of Mexican jungle and identifying a suspected unregistered site containing a plaza, an 8–12-meter oval pyramid, a possible reservoir, and three residential terraces. Mexico's National Institute of Anthropology and History confirmed in writing that the coordinates have no known registration record.

Factual Reconstruction

The episode began when the engineer, while traveling in Peru, discussed historical exploration cases with multiple models and then turned to public LiDAR data from the Mayan region. The agents converted LiDAR strips covering about 1,200 square kilometers into images, and the 48 agents reviewed them one by one, identifying geometric forms such as pyramids, platforms, and plazas while filtering out natural hills, sinkholes, and roads. The entire process used no specialized archaeological classification models and relied only on general-purpose models executing visual analysis according to precise guidelines.

The agents further cross-checked literature and archives and flagged that a major village had no registration record. The site is located at the edge of a cliff, with straight edges and sharp corners, and contains a level plaza, an 8–12-meter oval pyramid, connected elevated courtyards, and flat-bottomed depressions. The engineer then contacted Mexico's National Institute of Anthropology and History and submitted a report containing the coordinates. The institute replied confirming that no site was registered there.

Mechanism Breakdown

The core mechanism lies in parallel division of labor among multiple agents: some agents handled image conversion and preliminary screening, while others performed literature cross-validation and 3D point-cloud review to ensure that ground forms were not vegetation artifacts. The entire process, from data acquisition to report generation, closed the loop within 72 hours. During this period, the agents provided guidance on report writing, translation suggestions, and contact information.

This process relied on public LiDAR data and the visual understanding capabilities of existing general-purpose models, requiring no additional training or domain-specific tools. Through iterative dialogue, the agents gradually refined the task, from preliminary identification to generating 3D visualizations and speculations about historical context, forming a complete chain of discovery.

Industry Impact

This case shows that a swarm of general-purpose models can rapidly integrate public remote-sensing data and literature resources, lowering the barrier to entry for discoveries in specialized fields. Traditional archaeological work often requires long-term field surveys and specialized equipment, whereas this approach completes large-scale preliminary screening in a short time. It may change how resources are allocated, allowing researchers to focus on field verification earlier.

For research institutions, similar methods can be extended to other fields that use public remote-sensing data, accelerating the conversion from data to hypotheses. However, final confirmation still requires on-site verification by professional institutions; the agents' output only provides candidate locations.

Strategic Judgment

[Analysis] Without specialized models, a swarm of general-purpose agents has demonstrated the ability to process large-scale images and cross-validate literature. This may drive scientific research tools to evolve from single models to multi-agent collaboration. The future competitive landscape may shift toward who can more efficiently orchestrate general-purpose models to cover specific scientific workflows, rather than developing vertical domain-specific systems. Among stakeholders, institutions with large amounts of public data may gain greater advantages, while teams that traditionally rely on specialized tools need to reassess their workflows. In historical precedent, LiDAR technology itself has been used many times to discover Mayan sites. This case further shortens the time from data acquisition to preliminary reporting, but the field verification step remains irreplaceable.