One Amateur, Four AI Agents, and 72 Hours: The Accidental Archaeology Pipeline That Already Had a Map

A non-archaeologist used commodity AI agents and public NASA lidar to rediscover a Maya site in under 3 days — then learned professionals had already found it. The real story is the workflow, not the find.

On October 18, 2025, an unnamed technologist traveling in Peru had an idle thought about Hiram Bingham's 1911 discovery of Machu Picchu and wondered what modern tools could do. By October 21, Mexico's National Institute of Anthropology and History (INAH) had confirmed that no registered archaeological site existed at coordinates the technologist's AI agents had flagged in NASA lidar data. Three weeks later, the author discovered that Estrada-Belli et al. (2025) had already mapped the site — the agents had simply missed it during their literature review.

The workflow itself is the real product here. The author deployed 48 AI agents (Gemini, ChatGPT, Claude, Grok) to process roughly 1,200 km² of NASA G-LiHT lidar strips covering the Maya lowlands. The agents converted raw lidar into images and scanned them visually — no custom model, no classifier, just multimodal LLMs following a written guide to identify pyramids, platforms, plazas, and filter out natural features like sinkholes. The agents then cross-referenced flagged sites against published literature and INAH's archaeological register. The site itself, if the lidar interpretation holds, appears to be a small Classic-period Maya civic center (roughly AD 250–900): a leveled plaza, an oval pyramid 8–12 meters tall, a raised courtyard, a possible water reservoir, and surrounding platforms likely used as houses — all buried under 10–17 meters of jungle canopy on a hilltop cliff edge. Raw point-cloud analysis confirmed the shapes are ground-level, not canopy artifacts. The update is the most important paragraph. The agents failed at precisely the task that would have made this a genuine discovery: comprehensive literature search. Estrada-Belli et al. (2025) had already identified this site. INAH's register simply hadn't been updated yet. This is not a minor footnote — it's the central lesson. The pipeline is fast, cheap, and impressive at pattern recognition in geospatial data, but it broke down at the scholarly validation step, which is the difference between discovery and rediscovery. The generative potential is real but narrower than the author's framing suggests. The actual capability demonstrated is rapid, low-cost screening of public remote-sensing data by non-specialists. That's genuinely useful. Archaeologists have been doing exactly this with lidar since at least 2016 (Golden et al.) and scored a major find in 2024 (Auld-Thomas et al.). What's new is the labor cost: one person, no domain expertise, 72 hours. What's not new is the method, the data, or the conclusion. The author's closing claim — "we should probably lidar the whole planet and let the agents discover the rest" — captures both the promise and the blind spot. The bottleneck in archaeology was never image scanning; it was fieldwork, permits, cultural context, and the hard institutional work of registration and preservation. AI agents accelerate the easiest part of the pipeline while the hard parts remain human-speed. Mexico's federal archaeological protection law, which requires INAH authorization for any fieldwork, is a feature, not a bug. The story is a clean demonstration of a pattern we'll see more of: amateurs using AI to reach the frontier of a field's screening capabilities, then mistaking the screen for the science. The 72-hour timeline is impressive. The literature-search failure is instructive. The correction is honest. The revolution, if it comes, will look less like discovery and more like triage at scale.