My AI agents found an unmapped Maya site

Google’s software engineers deployed a fleet of 48 AI agents to comb through NASA’s lidar dataset of the Yucatán, and within 72 hours the system highlighted a previously unknown Maya pyramid. The claim is striking because the entire workflow, from raw point clouds to a flagged site, happened in under a quarter of a week.

Each agent was tasked with a narrow sub‑problem: cleaning noise, segmenting terrain, identifying geometric patterns typical of step‑pyramid architecture, and scoring candidate features. The agents shared intermediate results through a central coordinator, allowing the pipeline to scale horizontally without a single bottleneck. The design leveraged existing lidar processing libraries but wrapped them in autonomous loops that iterated until a confidence threshold was met.

The confidence score that triggered the alert came from a simple statistical model comparing the height, slope, and rectangular footprint of detected features against a reference set of known Maya structures. When a feature exceeded the model’s cutoff, the coordinator logged its GPS coordinates and generated a visual overlay for human review. This deterministic scoring kept the system from flooding archaeologists with false positives.

For developers, the episode shows that breaking a large data‑analysis task into many lightweight agents can accelerate discovery when compute is abundant. Building a similar swarm requires a robust task queue, clear data contracts between agents, and a transparent scoring function so that downstream experts can trust the output.

TakeawaySwarm‑style AI pipelines can turn massive remote‑sensing data into actionable findings quickly, but they still need expert validation.