My AI agents found an unmapped Maya site?
TL;DR Highlight
48 agents scanned lidar images and found a Maya site, but it was already mapped
Who Should Read
Developers wanting to run multiple LLM agents in parallel for image analysis
Core Mechanics
- Lidar turned into images for agents
- About 1,200 km² of lidar strips were converted to images.
- 48 agents went through them one by one looking for pyramids, platforms and plazas.
- They were told to filter out natural hills, sinkholes and roads.
- No special model, just looking at pictures
- No special model or classifier was used.
- Agents looked at the images directly and followed a guide.
- Agents also checked the literature
- The author asked them to check which villages were undiscovered via literature and archives.
- A few hours later one major unregistered village was reported.
- It had a plaza, an oval pyramid at least 8–12 m tall and a raised courtyard.
- Raw point cloud cross-check
- Agents checked the raw point cloud to confirm shapes were on the ground, not trees.
- Under 10–17 m of forest the ground traced the plaza, pyramid and courtyard.
- It was already mapped
- Estrada-Belli et al. 2025 includes the site.
- The agents missed it in the literature search, and only INAH's register lacked it.
- The author later confirmed this against the 2025 paper.
- Agents wrote and translated the report
- Agents explained how to contact INAH and drafted the report.
- It was translated to Spanish to improve reply odds, and the author trusted the agents since he doesn't speak Spanish.
Community Reactions
- Missed prior work criticized
- Commenters said it resembles an LLM finding a bug that was discussed years earlier.
- They worried LLMs convincingly overstate the significance of rediscoveries.
- Lidar data, not AI, was the credit
- The creator of a UK lidar map said it led to dozens of new finds like Roman villas and forts.
- They felt the story reflected lidar coverage and data availability more than AI.
- Questions on setup and cost
- Commenters asked how the 48 agents were coordinated and shared data.
- One noted the 72 hours was not agent working time.
- One said Gemini told them it was 834 images of 1200x1200 for 1,200 km².
- A similar experiment by a commenter
- A commenter explored NASA data with Claude for an hour and found what looked like a new asteroid pair, which had already been discovered this year.
- Verify before trusting
- Commenters said agent outputs must be rigorously checked.
- Others said the last 10% still needs domain expertise.
How to Apply
- Add a prior-work check to agent results
- Take one agent finding or bug report today and have a separate agent check docs or issue trackers for existing reports.
- Mark claims without source links as unverified.
- Batch-process large visual data as images in parallel
- Tile sensor, map or log visualizations into images and split them across agents.
- Give each agent a guide listing what to find and what to filter out.
- Verify candidates with raw data
- When image analysis yields candidates, have another agent confirm them against raw data.
- Escalate only confirmed candidates to humans.
Try It
# Paste into Claude or ChatGPT (simplified version of the method: image analysis plus prior-work check)
You are an analyst in {{field}}. In the attached images {{images}}, find regular structures that look like {{target}}.
Exclude: {{natural features to filter out}}
List each candidate in a table with location, evidence and confidence.
Then check via literature search whether each candidate is already known, and write 'unverified' if there is no source link.
Next to try: give the same images to a different model and compare results.Terminology
Related Papers
LLM assisted reverse engineering
Pointing an agent at binaries with skills and MD notes made years-long hacks feasible
Metrics-in-the-Loop: Complexity Feedback Reshapes Agent-Written Code
Injecting per-turn complexity reports into coding agents cut code erosion without changing correctness
Why is Gemini 4 Carbon taking so long?
Carbon took nearly an hour on bug fixes that Astra and Opus finished in minutes, mostly from extra unrequested work.
How the Claude Code team thinks about verification
E2E, property and differential tests plus verification rules let Claude run in parallel.
Running coding agents without handing them my credentials
Credentials stay on the laptop; agents run commands there via Touch ID approval
Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use
Reverse engineering shows Windows 11 logs which coding agents you use and reports it as optional diagnostics
The Silent Cost of Conversation History in LLM Reasoning
Keeping full conversation history dropped accuracy from 0.6 to 0.1 within 20 turns; a verified state summary prevented it.
Anthropic upgrades Claude Code Projects to run parallel AI workers
Claude Code Projects now has a coordinator that splits work into parallel threads and opens PRs.