The Market Agent — the brain of a place.

What the market agent is

If the advisor is one buyer's relationship, the market agent is one place's memory. It's the only thing in the system that scrapes, ingests, and reasons over the raw facts of a market — and the advisor reads its brain rather than ever touching a source itself.

It already does the thin slice (Zillow listings → Turso → per-home verdicts in Convex). This spec captures the parity rebuild of everything the old Delta agent did, on Flue: county records, a vector mind, and a wiki — all on the full storage stack every agent gets.

Where we are — as-built

Unlike the advisor, the market agent is real end-to-end for its current scope (flue/src/agents/market.ts + flue/src/market/{brightdata,publish,tools,store}.ts, 74 vitest specs).

The full storage stack

Every agent in HomeZero gets the same five places to keep things. The point of the market agent is that it's the one that uses all of them, hard. The advisor uses a lean slice of the same stack — shown alongside so the two specs read as one model.

Flue skills — same primitive

The market uses the identical skill model proven in the advisor spec: a skill is a SKILL.md in .agents/skills/, discovered from the workspace, loaded on demand via activate_skill (progressive disclosure), reusable across agents. Tools stay agent-level code; each tool is paired to a skill by convention — the tool is the hands, the skill is the know-how that teaches its use.

Self-improving skills · the metric skill

Self-improvement is built the same way (Flue has no native hook): a skill carries a metric from real signal and tunable state (params in Turso and/or its own SKILL.md body), and the nightly dream rewrites the playbook. The market is where this pattern is most obviously native — its missing "metric skill" is a self-improving skill:

// market's daily billboard / metric beat
const snap = brain.computeMarketMetrics()   // median asking, DOM, verdict-mix, honest-deal count
brain.writeMetrics(snap)                       // → skill_metrics (Turso)
const drift = compare(snap, lastWeek)         // what's moving in this market
fs.writeFile('.agents/skills/comps/SKILL.md', retune(body, drift))  // the comps playbook adapts

The same loop tightens the comps skill (do its comps agree with what actually sold?), the county-ingest skills (did a source's parse drift?), and the study cadence (which yearning is worth a pulse). Every market skill measures itself.

The brain — structured + vector

Structured spine — houses and their facts

A parcel-centred schema, jurisdiction-agnostic via an EAV facts table so a new county doesn't need new columns. Vectors are native libsql — F32_BLOB(n) + libsql_vector_idx (DiskANN) + vector_top_k — no Stoolap, no new dependency, just an ensureSchema add.

parcels( pin PRIMARY, geom, acres, owner_inst, … )        // the spine (ArcGIS + assessor)
parcel_facts( pin, key, value, source, observed_at )       // EAV — water, zoning, soil, CDL…
parcel_listings( pin, zpid )                              // bridge scalar listings → the spine
sales( pin, price, date, doc_ref )                         // real comps, not estimates
deeds · foreclosures · distress · permits · tax_history
listing_vectors( zpid, embedding F32_BLOB(1024) )         // the vector mind
brief_vectors(  slug, embedding F32_BLOB(1024) )
wiki_links( from_slug, to_slug, kind, weight )              // the comps graph (§06)
yearnings · skill_metrics · skill_params             // shared brain shape

The vector mind, used creatively

Pete's nudge — "a vector we can think of creatively." A few ways the market puts it to work beyond plain similarity:

• Comps that mean it. vector_top_k over listing_vectors + brief_vectors, then graph traversal of wiki_links — neighbours by meaning, cross-checked against real sales.
• Revealed substitution. Homes that buyers treat as alternatives cluster in vector space — the demand-side comps the MLS can't tell you.
• The shape of demand. Embed what's being zeroed, not just what's listed; the gap between the demand cluster and the supply cluster is the market's pressure.
• Under-priced outliers. A home whose price sits far from its vector neighbourhood's price band is either mispriced or mis-described — both worth a brief.
• Scout fuel. The advisor's scout ranks against these vectors (Phase 2) — "homes like the ones this buyer warmed to."

The wiki — a comps engine

The wiki is the market's edge that grows with every zero. It's not a separate store — it's the wiki_links edge table over the briefs, woven at publish time. Each new brief is linked into its neighbours (same street, same band, substitute, comp, shares-a-risk), so comps become a graph traversal seeded by a vector search, not a SQL WHERE. Built fresh — the old Delta agent never had it.

• Woven on publish_briefing — authoring a brief also writes its edges, so the graph is always current.
• Source for the advisor — a scout card's "why" can cite the wiki path that connected two homes.
• Self-improving — the comps skill's metric (do linked comps predict the sale?) tunes which edge kinds to trust.

The loop — five beats

Same primitive as the advisor (extend + scheduleEvery over DO alarms, cron as the floor). The market's daily study is bounded — one yearning per pulse on a study budget — which doubles as the source rate-limiter so it never hammers a county site.

The skill catalog

Market tool↔skill pairs. live = the tool exists today; new = parity rebuild. shared = same skill object as the advisor.

County data & egress

The egress question is settled (spike-verified): no county or state source needs a residential IP — that was a Zillow-only artifact. 8 of 9 sources are a direct datacenter Worker fetch(); the lone exception has a clean replacement. Residential egress stays scoped to Zillow.

So Phase 2 = direct fetch + an optional Cloudflare Browser Rendering binding for the one bot-walled source. No wells, no proxy. Each source is a pattern adapter sized to fit one study pulse.

Boundaries & the advisor handoff

• The market owns objective home knowledge — facts, briefs, comps, the wiki. The advisor owns the buyer relationship. Neither duplicates the other.
• The advisor reads the market's Turso (scout beat) — both DOs live in the one Worker, so the advisor opens a read client on the market's brain. The market never reaches into the advisor's.
• Two-author card. The clean split we want: the market authors the card's content (brief, blurb, full briefing); the advisor authors the question and the 👍/👎. The question is never the market's. (Today the "market half" is a render-time JOIN — making it an authored field is its own small schema task.)
• Convex is the meeting point — the thin shared projection both write into; the deep brains stay private.

Build status

The market agent is not in the current build. We ship the advisor first (1A–1D). This spec is the capture so the parity rebuild is ready to lift when its turn comes. When it does, the order roughly mirrors the advisor's:

Same skill model, same rhythm, the shared brain skill already authored in the advisor build. Nothing here is orphaned — it's queued.