Case Studies

How an AI Agent Helped Discover a Maya Site

Pietro Miotti’s Agent team helped locate a previously unmapped Maya site. The more revealing story is how the system connected online research, field investigation, and evidence preservation while human archaeologists retained authority over consequential decisions.

Font size

Pietro Miotti says his AI Agent team discovered a previously unmapped Maya site. The more revealing story is how it turned scattered clues into verifiable action in the field.

Original report: pietromiotti.com/latest/maya-site

An Agent finding a site is not independent archaeology

Start with the project workflow.

Miotti’s team gathered leads, then used them to plan a field expedition. On site, the team collected data before analyzing the material gathered in the field.

This chain involved four types of inputs and outputs:

  • Public data: historical documents, existing archaeological records, satellite imagery, or map information.
  • Search results: the Agent extracted locations, terrain features, and hypotheses to be verified from multiple sources.
  • Field actions: researchers traveled to candidate areas and recorded what they actually observed.
  • Analysis results: site photographs, measurements, and notes were fed back into the investigation.

Human involvement did not begin only with polishing the final answer.

An archaeological project must address permits, safety, navigation, sample collection, site protection, and professional judgment. An Agent can narrow the search area, organize maps, and cross-check information. But model confidence alone cannot determine whether a terrain feature indicates an artificial structure, whether a find must be reported, or whether a team can safely enter an area.

The more accurate account is that Miotti’s Agent team participated in a hybrid investigation process. The Agent provided search, planning, and analysis capabilities; human researchers supplied domain expertise, operational authorization, and final interpretation of the evidence.

That distinction matters.

A chatbot that lists ten locations where a site “might exist” has not completed an investigation. It has not converted candidate locations into an executable route, nor used field observations to disprove or support the original hypotheses.

Closing the loop from web search to the real world

The typical chatbot loop looks like this:

User asks a question → Model generates an answer → User decides whether it is satisfactory

This case involved a closer approximation to:

Gather leads
  ↓
Identify candidate locations and hypotheses to verify
  ↓
Plan a field route
  ↓
Survey the site and collect data
  ↓
Analyze field data
  ↓
Cross-check against maps and original records

The hard part is not merely knowing how to call tools.

First, tools do not produce the same kind of output. A map search returns coordinates, satellite imagery returns pixels, field notes record observations, and measurement tools produce numerical values. Each output has its own noise and error. If an Agent simply combines the results into prose, it can easily turn “appears to be” into “has been confirmed.”

Second, task state must persist across steps. After finding a lead, later planning needs to know where it came from, what constraints apply, and which parts remain unverified. A failure also cannot be silently skipped.

Third, the physical world produces unexpected results. Closed roads, inaccurate coordinates, dense vegetation, and changes in weather can invalidate an otherwise sound plan. An Agent needs iterative reasoning that adjusts the next step in response to feedback, not a single perfect chain of inference.

This is where multi-Agent collaboration becomes especially interesting in archaeological investigation. Roles can be narrowly defined: a retrieval Agent, map Agent, image-analysis Agent, planning Agent, and evidence-curation Agent. They might respectively produce candidate locations, coordinates, routes, observation records, and citation relationships.

More roles, however, do not automatically mean greater capability.

Without shared task state and a common evidence format, multiple Agents can easily produce little more than “several summaries citing one another.” Each Agent completes its own prompt, yet no one can answer the essential question: what is the original evidence behind this conclusion?

What the comments questioned after 908 points

The project reached the Hacker News front page with 908 points and 334 comments.

Discussion: news.ycombinator.com/item?id=50452455

The high score shows that the case was compelling. The comments also confronted a less convenient question: how much of the work did the AI actually perform?

1. Is the evidence chain complete?

People asked where the candidate locations came from, which maps and images the models relied on, how field data entered the analysis, and whether the relevant screenshots, coordinates, logs, and survey records could be made public.

That is the most important part of the case.

“Discovery of an unmapped site” is a result statement. The result can be brief, but the evidence chain should be extensive. Each step should ideally preserve its source, timestamp, original files, processing history, and points of human confirmation. Without those records, outside readers cannot determine whether the Agent led the investigation or merely organized work already completed by humans.

2. Where are the task boundaries?

Archaeology is not an ordinary web task.

A website can be retried and an API can be rate-limited. A lead incorrectly tagged as promising does not restore itself. Field movement, evidence collection, and site protection involve real-world risks. Any Agent’s autonomy must therefore be constrained by permissions and approvals.

“End-to-end Agent” does not mean “every decision is delegated to the Agent.” A more realistic engineering goal is to let the Agent work on evidence retrieval, repetitive analysis, and route proposals while reserving irreversible actions for people.

3. How should autonomy be defined?

If a system generates candidate locations, humans drive to them, and then confirm the site, how Agentic is the result?

Three concepts are often conflated: the degree of automation, an Agent’s autonomy, and attribution of the outcome. The more steps a system performs automatically, the more judgment authority it has; nor does human assistance with transportation, navigation, or key observations mean the Agent contributed nothing.

A more productive discussion would break the project into measurable steps: How many candidates did the system propose? How many did it eliminate? How many did humans reject? How much new evidence did the field team collect? How much of the final positioning came from machine judgment?

4. Who should receive attribution?

Attribution is not simply a contest between “multiple Agents” and “humans.”

Archaeological records typically include many contributors: the person who identified the lead, the person who supplied historical material, the field researchers, the image classifiers, and the people who maintained the data pipeline. An Agent’s contribution should be tied to specific outputs rather than summarized vaguely as “AI discovered.”

That approach is more cumbersome, but it is better suited to auditing.

The real engineering cost of multi-Agent collaboration

The appeal of multi-Agent systems is straightforward: break a large problem into smaller parts so that each role handles only one segment.

In real engineering, decomposition itself creates cost.

Task decomposition

If the split is too coarse, one Agent must handle maps, historical records, imagery, and routes at once, making errors difficult to isolate. If it is too fine, the system produces excessive handoffs.

A more useful decomposition is based on verifiable artifacts rather than names inspired by human personalities:

  • Location retrieval: produce candidate locations with sources.
  • Image inspection: identify areas of interest and uncertainty.
  • Route planning: produce an executable itinerary and alternatives.
  • Field recording: preserve original files, timestamps, and positions.
  • Evidence curation: establish “claim–evidence–source” relationships.

Each role should have explicit completion criteria.

Communication

Agents should communicate through structured objects rather than long passages of natural language.

For example:

{
  "candidate_id": "site-017",
  "coordinates": {
    "latitude": null,
    "longitude": null
  },
  "claim": "Signs of artificial structures await field confirmation",
  "evidence": [],
  "confidence": "low",
  "requires_human_review": true
}

Messages like this can be traced to their sources and validated by later steps. Natural-language summaries are useful for people, but they are poorly suited to storing state.

Failure and retries

A failed search, a rate-limited map service, an unclear image, and an inaccessible site each require a different response.

None of them should be reduced to a generic “try again.” A retrieval task can change keywords or sources; route planning can calculate an alternative destination; a visual judgment can be downgraded for manual review; and an action that requires entering a site must be stopped and escalated.

This exposes a common weakness in multi-Agent systems: they tend to search for a recovery path that sounds plausible rather than acknowledge that a reliable alternative does not exist.

Evidence preservation

Evidence preservation should include at least:

  • Input sources and access times.
  • Original maps, images, photographs, and measurement data.
  • Intermediate Agent outputs.
  • Tool-call records.
  • Records of human modifications and approvals.
  • The evidence chain supporting each final claim.

The more an Agent can do in the physical world, the higher the cost of auditing it.

Otherwise, the system leaves behind only its final conclusion, and outsiders can neither reconstruct the process nor distinguish model judgment, retrieval errors, and human experience.

Multi-Agent systems can add value only when tool interfaces are clear and evaluation criteria are explicit. Without those conditions, the approach merely divides one complex act of reasoning into several rounds of mutual guesswork.

Editorial perspective

The significance of this case is not that “AI found treasure.” Archaeological discovery requires professional credentials, long-term fieldwork, and responsibility for conservation. An Agent’s participation in search and evidence organization does not turn the story into one of independent machine discovery.

My view is that the value of Agents is beginning to shift from content production toward process execution. Web retrieval is no longer scarce. The new engineering challenge appears when maps, imagery, research records, and field observations are connected into an inspectable loop: How should permissions be set? How should evidence be preserved? When should a process stop? Where and how should humans intervene? For corporate investigations, operational inspections, and supply-chain verification, these questions are harder than producing a few additional pages of reporting and much closer to real production work.

This case differed from earlier demonstrations: the endpoint was not a polished answer but a previously unmapped site. Even so, the public material does not provide enough detail to confirm the complete method for reproducing the result. There is insufficient information about how the Agents were divided, which models were used, and which steps were completed by humans. The case merits attention, but its conclusions should remain conditional.

Source: https://pietromiotti.com/latest/maya-site