The plan for spinning HomeZero up almost anywhere in the country — what we stand up the moment someone signs on, what waits, and the one genuinely hard question underneath all of it.
Every agent is bound to a durable, real-world thing — and the binding is the design. It says who follows you, who knows the place, and who keeps the house honest.
Your single point of contact, for life. Follows you across every home you'll ever look at — and the one you'll live in. Knows you, not the market.
stands alwaysThe local expert for one market. Holds the public-record spine, knows what anything is worth, and analyzes a home the moment a buyer zeros it. Shared by everyone looking there.
stands alwaysBound to the building itself, not whoever owns it. The home's own resident intelligence. Born only when a buyer gets serious enough to go scan it.
born on a scanThe binding is also the governance rule: house-facts travel with the twin; person-facts stay private with the advisor.
We used to reach for a new specialist cell every time a buyer needed something (water rights, land, financing). We've stopped. A new capability is a skill the two standing agents carry — not a new agent. Fewer agents is dramatically more manageable; the real cost of a cell isn't compute, it's coordination.
Something only becomes its own cell when it passes a single bar:
“Does it dream?”
Our cells consolidate memory nightly — they dream. If a capability has nothing worth sleeping on — no evolving model that gets better over time — it's a skill. A water lookup is a skill. A water desk that's tracked a whole basin all season might be an agent. Even then: build it as a skill first and let it earn promotion.
The advisor spins up instantly anywhere — it's a person-bound chat agent with no local-data dependency. The market agent is the only heavy birth, and it's only instant where our scrapers already bind. Everywhere else it comes up partial and honest, and the gaps become the work queue.
Location is required — fuzzy is fine. “North County San Diego,” “out near Paonia,” “somewhere on Long Island.”
Mother turns the fuzzy phrase into a market manifest — a real boundary, which is almost never a county. One market? Or does this split? (this is the hard part →)
Instant, cheap, works on day one anywhere. Your lifelong point of contact comes online.
If a market cell already covers the place, attach to it. If not, birth one — the only expensive birth in the system.
“Can I actually work here?” It probes whether its scrapers bind to this jurisdiction's records and emits a green/flagged report. Flags = the backfill queue. We go live partial, never pretend a half-sourced market is covered.
Install the Chrome extension and start starring (“zeroing”) homes on Zillow. That's the demand signal everything downstream runs on.
Two cells stand up at onboarding — advisor + market. That's the entire standing team. Everything else is a skill they carry.
This is the question the whole product rests on. A market agent obviously can't hold millions of homes — there's a technical ceiling and a logical one. But a county is the wrong unit, and so is a metro. So what bounds a market?
The “millions of homes” fear assumes the market agent holds every home. It doesn't. Because we're demand-driven, it only deep-analyzes the homes buyers actually zero, and watches those. The heavy parcel/record data is a shared layer keyed to the record-keeper — queried, not held. So what really sizes a market is the comp model + the watch load: how rich a “what's anything worth here” model one cell can keep current, plus how many live homes it's tracking. Call it the attention budget — roughly a few hundred to a couple thousand active listings, not tens of thousands.
Do the same buyers actually weigh these homes against each other? Comps, schools, commutes, the feel of the place. Don't split a real market.
Can one cell hold it in its head and keep it current and still sound like a local? Don't let it sprawl.
When a market outgrows the budget, it splits. When one goes idle, it merges with a neighbor.
Coastal North County (beach, premium, walkable) and inland (big lots, wildfire backcountry) are different buyer worlds inside one assessor.
Paonia–Hotchkiss–Crawford is one market — orchards, water, off-grid, remote work — that crosses three county lines.
RI abolished county government in 1842. The town does assessment and deeds. The whole Colorado water layer is irrelevant; new local layers replace it.
“Recording jurisdiction” (who publishes the records — county, town, parish) is plumbing, shared. The market is the agent's identity, sized by attention, not by an administrative line.
This is the move that dissolves the whole problem: spin up a skinny market agent fast on a rough boundary, then let the boundary sharpen over time. The edge is a living hypothesis, seeded cheap and proven by demand — across three timescales.
Mother guesses a provisional boundary — one Zillow region, a few ZIPs, a school district — plus a confidence. Good enough to start advising immediately.
A one-time, cheap pass over free national data — not a listing scrape. Price-series clustering, school zones, commute sheds, migration. Proposes a real manifest.
The best signal no dataset has: which homes do the same buyers zero? Co-consideration is the market. The agent reshapes its boundary — splits, merges — from real behavior.
The strongest purely-structural signal, and it's free and national: pull each ZIP's home-value time series and see whose curves move together. ZIPs that rise and dip in lockstep are one market; a curve that goes its own way is an edge.
The other v1 layers stack on top of price clustering: school attendance zones (buyers substitute hard within a district, rarely across), commute and labor sheds (where people live-vs-work bounds “where else would I live”), migration flows (revealed substitution at the macro scale), and plain physical geography — a mountain pass, a coastline, a freeway is often the literal wall. Cluster them, propose a manifest, flag “this is really three markets.”
Almost never at the moment it's zeroed. A casual Zillow star is the weakest possible signal and most never convert — birthing a cell for each would be madness. The market agent carries every zeroed home cheaply. A home earns its own cell only when the buyer goes to touch it in person.
Stars a listing on Zillow via the extension. A demand signal enters the system.
One Zillow detail fetch + the public-record layers → a full Brief. The advisor turns it into a take that means something to this buyer.
The market agent watches price, status, and distress until the home sells or delists. The buyer keeps looking. Could be one home or fifty — all cheap.
This is the birth. The buyer goes to walk it, photograph it, record it — collecting new ground-truth no scraper can get. Now the home twin is born: it ingests the scan and becomes the home's resident intelligence.
We just cleared the specialist cells and a duplicate. What's left is the two-agent spine plus the Delta County reference build we keep around to harvest skills from. The naming rule: everything in this project starts with zero-; anything else belongs to the platform or another project.
| Cell | What it is | Status |
|---|---|---|
| zero-mother | Births the team; brain being repointed from “what specialists?” to “what place?” | keep · redo |
| zero-delta-co-market | The productized market agent — the template we'll stamp out elsewhere | keep · redo |
| delta-market | The rich Delta prototype (≈50 data layers) — reference; harvest its skills | keep · redo |
| zero-abstractor | National parcel-resolver (listing ↔ parcel join) | keep |
| zero-pulse | Heartbeat / scheduled wakeups for the zero fleet | keep |
| zero-delta-co-land | Specialty cell — the path we're abandoning | killed |
| zero-delta-co-water-rights | Specialty cell — the path we're abandoning | killed |
| abstractor | Duplicate of zero-abstractor | killed |
| mother · pulse | Platform / system cells (run the whole fleet) | not ours |
| tend-advisor-pete | A different project entirely (Tend) | other project |
How green does a market's readiness report need to be before we let a buyer onboard into it? My lean: go partial, advisor works day one, flags drive backfill.
What's the actual ceiling — active listings? watched homes? — that triggers a split? We need a real number, learned from Delta.
Right now there are two market cells for one place. The redo should converge to a single canonical template that births anywhere.
Today the twin is short-lived. The vision: it persists with the home and conveys to the next owner. When does that scope become real?