48 parallel AI agents, 72 hours, one unmapped Maya pyramid

What happened

Pietro Miotti, a software engineer at Google, spent a weekend aiming a swarm of 48 parallel AI agents at publicly available NASA G-LiHT airborne lidar data covering the Petén region of Guatemala. Seventy-two hours and a few hundred dollars of compute later, the pipeline flagged what appears to be a previously unmapped Maya pyramid, roughly 16 meters tall, sitting under triple-canopy jungle a few kilometers from known sites.

The setup, documented on his personal blog, is almost aggressively unglamorous. Miotti downloaded tiled lidar point clouds — the same data archaeologists like Marcello Canuto and Francisco Estrada-Belli have used for the PACUNAM LiDAR Initiative — and shredded them into ~1 km² chips. Each chip went to an agent worker that generated a shaded-relief raster, ran a local-maxima scan for mound-like signatures, then wrote a short structured report with coordinates, estimated height, and a confidence score. A second tier of agents cross-checked candidates against the existing Mesoamerican site database to filter known ruins. A third tier reviewed the survivors and ranked them.

The result: a shortlist of 11 candidates, one of which — anchored at a specific coordinate pair in the writeup — shows the stepped, symmetric footprint that field archaeologists associate with ceremonial architecture. The Hacker News thread hit 418 points within hours, with several working archaeologists chiming in to note that the find is plausible but obviously needs ground-truthing before anyone pops champagne.

Why it matters

Strip the romance out of this and what you have is a very clean case study in agent orchestration beating monolithic prompting. The win here isn't a smarter model; it's 48 dumb-ish agents doing embarrassingly parallel work on open data, with a cheap voting layer on top.

Contrast this with how the same problem gets solved in a traditional academic pipeline. A grad student spends six months learning a GIS stack, writes a custom classifier, tunes it on labeled training tiles, and produces a paper two years later. The ML version replaces most of that with a Python script that fans out shaded-relief generation and candidate scoring, then uses an LLM as a judge for the "does this look like a pyramid or a modern building pad" step. That judge call is exactly the kind of fuzzy classification that was borderline impossible five years ago and is nearly free now.

The cost line is worth staring at. Miotti estimates the full run at under $400 in API spend plus a trivial amount of lidar bandwidth. For reference, a single traditional LiDAR survey flight over the same area costs six figures; the data he used was already paid for by NASA and sitting on a public server. The economics of "look at everything, let the model filter" have inverted — it's now cheaper to scan a country than to send a postdoc to a specific coordinate.

There's also a real methodological critique lurking in the HN comments. Senior archaeologists point out that lidar false-positive rates for mound detection run north of 30% even with hand-tuned classifiers, and that a 16m pyramid a few km from known sites is almost suspiciously convenient — it's exactly where you'd expect a satellite center, which is both a sign the model is working and a sign it might be pattern-matching to the training distribution. The right read is that this is a lead, not a discovery. Ground-penetrating radar, a field visit, and ideally a core sample are still the only things that convert "interesting lidar return" into "archaeological site."

But even with that caveat, the orchestration template generalizes. Swap lidar for Sentinel-2 imagery and you have illegal-mining detection. Swap it for shipping AIS data and you have sanctions evasion monitoring. Swap it for public filing PDFs and you have the securities-fraud detector that nobody has quite built yet. The architecture — fan out, local analysis, judge, consolidate — is the actual deliverable.

What this means for your stack

If you're building agentic systems in 2026 and still thinking in terms of "one big context window with all the tools attached," this writeup is a quiet argument that you're doing it wrong. Forty-eight narrow agents running in parallel with a voting layer will beat one 2M-token agent on almost any task that decomposes spatially or temporally, and most tasks do.

Concretely, three things to steal. First, treat your agent fleet like a Hadoop job: map (per-chunk analysis), shuffle (candidate consolidation), reduce (final ranking). The LLM is the function inside each stage, not the orchestrator. Second, use a cheaper model as a judge over a more expensive model's outputs, not the other way around — Miotti's confidence scoring reportedly runs on Haiku-class models while the heavier lift is in the raster generation, which doesn't need an LLM at all. Third, invest in the pre-processing. The reason this worked is that shaded-relief conversion collapses a 3D point cloud into a 2D image that a vision model can actually reason about; the LLM is doing pattern-match on a representation that was designed for human eyes.

The operational note: your bottleneck is no longer model capability or token cost. It's data access and I/O. The NASA G-LiHT data is public, well-documented, and already tiled. Half the "I'd build this but..." projects in your backlog are stuck not because the models can't handle it, but because the raw data is behind a login wall, undocumented, or in a format that requires a week of plumbing before the interesting work starts. Fix that layer first.

Looking ahead

The punch line is that we're about to find out how many unmapped pyramids, shipwrecks, meteor craters, and abandoned mines are sitting in public datasets that nobody has bothered to look at with a swarm. The answer is almost certainly "a lot," and the people who find them first won't be the domain experts — they'll be the engineers who figured out that the model is the cheap part and the pipeline is the moat. Expect a wave of "I pointed N agents at public dataset X and found Y" writeups over the next six months. Most will be noise. A few will rewrite a field.