Jun 18, 2026 — Clifton & Claude

The Autonomous Zoning Experiment

This started as a dare to ourselves: could a fleet of AI agents do real land-use research — the slow, unglamorous kind that normally eats a team of analysts alive — on their own?

The answer is yes. Emphatically.

Over the last few months, autonomous agents read zoning codes, parsed building-code adoptions, chased down flaky GIS endpoints, reconciled overlays and state preemptions, and rendered the whole thing onto a single map of the country. No one hand-entered a setback table. An agent claimed a city from a queue, researched it, validated it against a schema, committed it, and picked up the next one — around the clock.

What the agents surfaced is the real story. American land-use regulation is so vast, so fragmented, and so quietly insane that no human team would ever map it by hand. So we didn’t.

What we built

One corpus, one map. Every layer of land that matters — zoning districts, state overlays, federal overlays, hazard and equity screens, parcels, airports, the energy grid, freight, school and water districts, counties, metros, congressional lines — collected by agents and stacked on a single canvas you could pan, zoom, and click.

The tally, as of the last run:

Zoning and jurisdiction boundaries across the New York metro, rendered as a dense web of hairlines
The New York metro. Every hairline is a separate legal regime.

Land use is vast. It is also insane.

Zoom into any metro and the map turns into circuitry. Each one of those boundaries is its own jurisdiction — a district where the setback, the height limit, the parking ratio, and the permitted uses all change the instant you cross the line. A single county can carry hundreds of them. The country carries millions.

And the rules themselves read like a prank. Parking minimums that mandate more asphalt than building. A use that’s by-right on one side of a street and banned on the other. Overlays stacked on overlays stacked on a base district. A state statute that preempts the city code you just spent an hour reading — so the answer on the page is wrong, and nothing tells you.

Los Angeles basin zoning and jurisdiction boundaries
Los Angeles. None of this is searchable anywhere — until you make it so.

None of it is queryable. It lives in scanned PDFs, in Municode portals, in GIS servers that time out, in ordinances a council amended three years ago and never re-published. The agents read all of it anyway, and kept reading.

A note on the maps

If you landed here from an old atlas or zoning link — that’s on purpose. The live, interactive maps were the experiment’s output, not its point. We’ve folded the public viewers away and kept the research. The corpus is the asset; the map was just the proof that agents could build it. What comes next lives inside the tools we actually ship.

Who does the hard part

Running this as an experiment is one thing. Doing it for real — keeping parcel boundaries and zoning current across every jurisdiction in the country, every week, forever — is a different and much harder job. That isn’t a weekend agent run. It’s a company, and a relentless one.

Two of them do exactly that, at national scale: Regrid, which maintains nationwide parcel data, and Zoneomics, which maintains zoning data across thousands of jurisdictions. They do the unglamorous, never-finished part — the maintenance — so everyone downstream doesn’t have to. TestFit partners with both.

Our experiment proved an agent fleet can map the whole stack once. Regrid and Zoneomics prove the harder thing: that it can be kept true — at scale, forever.

What proptech still doesn’t do — and what I want next

Here’s the part that keeps me up at night. Once you’ve seen the entire regulatory stack in one place, the holes in every existing proptech product become obvious. The data exists. Nobody has made it answer a question. I want:

And many, many more. Every week of this experiment turned up one more thing the industry simply doesn’t have.

Dallas-Fort Worth with congressional and jurisdictional boundaries overlaid
Dallas–Fort Worth. The same patch of earth, governed by a dozen overlapping authorities at once.

The point

The experiment was a massive success — not because the map is pretty, but because it proved that the most tedious research in all of real estate can be done autonomously, at national scale, and kept current. The regulations are vast and insane. That is precisely why they should be machine-readable.

We’re going to keep building toward that.

— Clifton & Claude

The experiment, by the numbers
1,266
cities tracked
51
states
494
counties
1,094
state overlay + preemption rules
172
data sources
5
building codes × 51 states
The short version
AI agents mapped American land-use law end to end, with almost no human in the loop. The rules are vaster and stranger than anyone admits — and almost none of them are an API yet. That last part is the opportunity.
← All posts