Skip to content

Google SWE 48 AI Agents NASA Lidar Maya: How to Replicate

Google SWE points 48 AI agents at NASA Lidar, finds unmapped Maya pyramid in 72h - here's the #1 mistake and exact agent workflow you can run on public data.

6 min readBeginner

The #1 Mistake People Make With the 48 AI Agents NASA Lidar Maya Story

Most people paste one mega-prompt into ChatGPT: “Scan this lidar and find pyramids.” That dies fast. Single-turn vision chats invent geometry, skip archives, and never produce paperwork a heritage agency will answer.

What held up for Pietro Miotti was the opposite. A swarm of ordinary vision agents, one shared visual checklist, tiled images in parallel, forced literature passes, raw point-cloud checks, then a human-edited Spanish report. No custom detector. Compute bought patience.

Headlines split the room: agents ate archaeology, or they only hit an INAH register gap on a site already in Estrada-Belli 2025. Both can sit on the same mound. Concrete loop below – on public data, without Google infra.

Quick Context (as of Nov 2026)

Miotti – AI researcher at Google’s Paradigms of Intelligence lab in Zurich, not a field archaeologist – was in Peru when the Bingham prompt hit. Gemini, ChatGPT, Claude, and Grok all pointed at lidar. Peru strips were thin, so the run pivoted to Maya lowlands and NASA G-LiHT (~1,200 km² of 2013 airborne data). Overnight, 48 agents turned strips into bare-earth tiles and scored platforms, plazas, mounds. One hilltop cleared the first archive pass. He cleaned the report, AI-translated it, mailed INAH. Next-day reply: nothing registered at those coordinates. Idea-to-reply under 72 hours. Full diary: pietromiotti.com.

Update already live on his post: same features sit in Estrada-Belli et al. 2025 (ScienceDirect). Agents missed the paper. Still a fast register-gap finder – not a first discovery.

Funny echo: Bingham needed porters, letters, and years of rumor. Here the bottleneck was vision tokens and whether anyone forced a Scholar pass before celebrating.

Hands-On: Replicate the Swarm on NASA G-LiHT

You do not need 48 instances. Four to eight parallel agents with vision + web tools (Claude Projects, ChatGPT with browsing/code, Gemini Gems, or open stacks) is enough to rank candidates and draft a report.

Step 1 – Grab and prep the data

Start at the G-LiHT Data Center or NASA Earthdata. Filter Mexico / Yucatan / Campeche (2013 AMIGA-Carb-era flights). Pull bare-earth DTMs, CHMs, or LAS/LAZ. Prefer OpenTopography reprocessings when they exist – original NASA ground calls often break under tropical canopy.

Tile it yourself or let an agent drive:

# rasterio + matplotlib - agent can edit paths/DPI
import rasterio
import matplotlib.pyplot as plt
with rasterio.open("dtm_strip.tif") as src:
 data = src.read(1)
 plt.imshow(data, cmap="terrain")
 plt.savefig("tile_001.png", dpi=150)

Target 200-500 m tiles, hillshade or openness + height color. Keep flight ID and coords glued to every file.

Step 2 – Write the guide, launch parallel agents

Same system text on every worker:

You are a patient remote-sensing archaeologist. For each bare-earth lidar image: 1. Flag straight edges, leveled platforms, plazas, mounds ≥5 m, courtyards, possible aguadas. 2. Reject natural hills, sinkholes, modern roads, drainage. 3. Score confidence 1-5 and say why. 4. JSON only: {tile_id, features: […], coords_approx, confidence, reject_reason}. Never invent coordinates.

One tile (or tiny batch) per agent. Collect JSON. A coordinator merges, dedupes, ranks top ~20 by confidence and geometric regularity.

Then lock the literature gate: search Scholar, JSTOR abstracts, INAH catalogs, Estrada-Belli and known site lists. Keep only zero-clear-match rows. Skip this and you replay the 2025 miss.

Nobody published the exact 48-agent wiring – which model ran the swarm, how tiles were sharded, token burn. Community write-ups (HN thread and follow-ons) flag that gap. Treat 4-8 agents + the guide above as a reverse-engineered slice, not a bit-for-bit clone.

Step 3 – Raw points, then the report

The catch is canopy. Forest on these strips runs about 10-17 m. Open survivors in CloudCompare – or code-interpreter vertical slices – and confirm brown ground returns trace the shape, not green vegetation noise. Fake mounds love misclassified returns.

Last prompt: one-page INAH-style note – description, figures, methods, offer to support a permitted survey. Translate to Spanish. You edit Spanish and the confidential coords annex yourself. Send only with real evidence. Mexican sites are federal; fieldwork needs INAH authorization.

That is the loop. Miotti slept through most of the agent hours.

Common Pitfalls That Kill These Runs

  • Skipping raw point-cloud slices under dense canopy → false mounds and missed platforms.
  • Stopping at “looks artificial” without forced Scholar/INAH/Estrada-Belli search.
  • Publishing coordinates. Register gap ≠ open season.
  • Low-res optical satellite instead of lidar DTM/hillshade. Trees win.
  • One giant context window. Vision tokens explode; tile and parallelize.

What Results Look Like (and What They Don’t)

Original top hit: cliff-edge hilltop, leveled plaza, 8-12 m oval mound read as pyramid, raised courtyard, flat-bottom hollow (possible aguada), three platforms. Classic-period guess (AD 250-900) was speculation. 3D exaggeration made it look civic. INAH had no register row. Boots still required.

One night of agents plus a human edit beat months of manual first-pass strip scanning. Cost is mostly vision tokens and your polish time. Your runs will mostly return known sites or false positives. Fine – ranked shortlist + draft paperwork is the product.

So where does “unmapped” actually live – the federal register, or the journal record? That gap is the whole story.

When NOT to Bother

No lidar for your box on the map? Stop. No legal path to report? Stop. Want a press-ready “new city” without experts? Wrong tool. Pure single-model chat is not this method. Swarm + guide + cross-check is. Coding agents already like this pattern; science-archive swarms are still noisy.

Side threads if you keep going: multi-agent orchestration, public Earth-observation pipes, and why environmental lidar (Cook et al. 2013 and later archaeology re-uses) keeps feeding site detection.

FAQ

Was it really an unmapped Maya pyramid found by a Google SWE in 72 hours?

No. Researcher, not SWE-as-job-title hype. Lidar mound, possible pyramid. INAH register empty; Estrada-Belli 2025 already had the features. 72 h = idea to agency reply.

Can I run this exact 48-agent setup today without a research budget?

Not exact – model mix, tiling, and cost were never fully documented. Smaller swarm yes: free G-LiHT tiles, 4-8 vision agents, the checklist above, a literature tool, raw-point validation. Start with 50-100 tiles. You will not match overnight scale; you will get a ranked list. Tokens are the leash.

Do I need archaeology training or a custom ML model?

Miotti had neither. Agents followed a plain-language visual list and did archive grunt work. The failure mode is soft discipline: skipping cross-checks, leaking coordinates, or treating a confidence-4 JSON row like a confirmed temple. Outsider willingness got the run started. Verification still sits with INAH and field specialists – outsider naivety is not a permit.

Grab one Mexico G-LiHT strip tonight. Write the guide. Launch four parallel sessions. Post the false-positive rate you actually measured – that number beats another hype quote.