The platform runs the realtor's buyer pipeline as a redesigned domain — not a chat assistant bolted onto a CRM. Each realtor gets a private container with eight composable atomic agents (matching, drafting, scoring, outreach, contacts, deals, outcomes, audit) that execute the same end-to-end workflow McKinsey's analysts identified as the highest-leverage redesign target in real estate. Average match time per listing is under 2 seconds. The realtor owns every byte of trace data, every learned scoring weight, and every fine-tune of their outreach voice.
TL;DR
- Architecture: maps directly to the McKinsey 5-layer reference architecture for agentic AI in real estate (March 2026)
- Composition: 8 atomic agents (MCP tools), not one heroic chatbot
- Inference: Claude Haiku for agent reasoning + a real-estate fine-tuned local model (
gemma4-realty:v1) for embeddings and extraction - Isolation: per-tenant containers; one realtor's data never trains another realtor's matches
- Speed: < 2 seconds per match; 24/7 after-hours buyer engagement
- Compliance: state-aware NAR-settlement-compliant buyer representation workflow built in
Why "atomic agents" instead of an AI Copilot?
McKinsey is explicit on this: "The winning operating models will not be built around a single heroic agent that tries to do everything. They will be built from atomic agents that do a small thing well, with clear boundaries."
A "Copilot" or "AI Assistant" is the heroic-agent pattern — one big model trying to do every step in a conversation. It works for demos and breaks at scale. The atomic-agent pattern is eight small composable tools, each with a single capability, an explicit allowlist, and a recorded audit trail. The realtor's skills compose them into workflows. Adding a new vertical means adding new atomic agents to the library — not retraining a monolithic chatbot.
The eight atomic agents shipped today:
| Atomic agent | What it does | When it runs |
|---|---|---|
cma-advisor | Pulls 3–6 comps for a subject property, returns {fair, low, high} with sourcing | On-demand for any listing |
listings | CRUD over the realtor's tracked listings | Continuous |
contacts | CRUD over buyer + seller contacts | Continuous |
deals | Pipeline state machine | Continuous |
inbox | Drafts outreach in 2–3 voice variants per send | When a match fires |
outcomes | Logs verdict (OFFER / FAIR / SKIP) → tunes scoring weights | After every recommendation |
alerts | Notification queue for the realtor | Continuous |
audit | Tenant-scoped query over the trace data the realtor owns | On-demand |
How does the platform map to McKinsey's 5-layer reference architecture?
McKinsey's March 2026 article identifies five layers every agentic AI deployment needs. The platform implements each one explicitly:
| McKinsey layer | What it does | How it's built here |
|---|---|---|
| Factual | Clean data, consistent identifiers, source of truth | Supabase tables with row-level security; property_catalog with regex-seed + hand-override + GIN aliases; pgvector embeddings produced by gemma4-realty:v1 |
| Orchestration | Plans, routes work, identifies stop points where humans must decide | Skills (realtor-buyer-match, realtor-context) running on Claude Haiku, with industry.config.ts declaring the allowed tools per workflow |
| Action | Executes work inside systems of record | The 8 atomic agents (MCP tools) above, scoped by per-tenant JWT, rate-limited by cost-guard |
| Control | Permissions, audit trails, monitoring | agent_audit_log records every tool call; cost-guard trips on Anthropic spend or gemma4 RPM; Supabase RLS enforces tenant isolation at the database level |
| Building-block | Library of small, reusable, composable agent blocks | Every MCP tool is one block. Vertical packs declare which blocks compose into which skills. New verticals reuse the library. |
This isn't aspirational — it's the architecture shipped today. McKinsey's diagram and ours overlay 1-for-1.
What does "automate the steps, protect the thoughts" mean for a realtor?
McKinsey's productivity sidebar splits every domain workflow into two ingredients:
- Steps: repeatable tasks that benefit from speed and consistency. Pull context, draft a message, route, schedule, log. Automate aggressively.
- Thoughts: judgment calls that require taste, discretion, or trust. Which buyer to call first. Which voice to send. Whether to disclose the off-MLS pocket listing. Protect deliberately.
In the realtor's day, here is what falls on each side:
| Step (the platform automates) | Thought (the realtor decides) |
|---|---|
| Pull listing details from MLS, Zillow, Redfin | Which buyer is most likely to actually move on this property |
| Score every active buyer against the new listing | Whether to lead with the price reduction or the school district |
| Surface the top 5 with confidence breakdown | Which DraftCard variant sounds most like you |
| Draft 2–3 outreach variants per buyer | When to call vs text vs email |
| Log the outcome verdict | Whether to disclose the off-MLS option |
| Tune the scoring weights from the verdict | Whether to make the offer call today or wait for the inspection |
The realtor never loses a thought decision to the platform. The platform never makes the realtor wait for a step.
Why a per-tenant container instead of a shared model?
Two reasons, one technical and one strategic.
Technical: every realtor has a different farm area, a different buyer pipeline, a different outreach voice. A shared model averages all of these into a generic experience. A per-tenant container learns the realtor's actual preferences from their outcome verdicts and tunes scoring weights to match. Match quality compounds with use.
Strategic: McKinsey raises the load-bearing question — "Who owns the learning loop — the owner, the property manager, the software vendor, or the services provider?" In every competing tool, the vendor owns the loop. The realtor's preferences become the vendor's IP. In this architecture, the realtor owns the loop. Their match weights, their outreach voice fine-tune, their outcome history — they take it with them if they leave.
What about the new buyer-rep agreement rules from the NAR settlement?
Buyer representation agreements are now mandatory before showing in most states, with state-specific divergence (Texas requires a written agreement before substantive action effective 2026; Alabama and Mississippi allow touring before signing). The platform ships state-aware buyer-rep agreement templates as part of the realty vertical pack, with auto-updates when rules change and commission disclosure language pre-filled. Compliance is built into the same workflow that drafts the buyer outreach — not a separate tool to remember.
What does the realtor pilot cost?
The pilot is free for 14 days. Invite-only. No credit card required. Production pricing is published when the pilot exits.
How does the platform handle privacy and PII?
Each tenant runs in an isolated container with its own Supabase row-level-security scope. PII is redacted before any data leaves the container for embedding generation on the local data layer (gemma4-realty:v1 runs on a private host, not a third-party API). Anthropic Claude Haiku receives only the structured prompts the skill explicitly assembles — never the raw buyer financial data. Every tool call is recorded in agent_audit_log and the realtor can export the full trace at any time.
Where does the platform stop, and where does the realtor take over?
McKinsey calls these "stop points" — the moments where confidence is low or risk is high and the system explicitly hands off. The platform's stop points:
- Always: sending the actual outreach message. The platform drafts; the realtor sends.
- Always: making the offer call. The platform scores; the realtor decides.
- Always: signing the buyer representation agreement. The platform fills the template; the realtor and the buyer sign.
- When confidence is low: the platform flags low-confidence matches and surfaces the breakdown rather than auto-recommending.
- When the cost-guard trips: the platform halts the active session and surfaces the incident before any further work runs.
The realtor stays in control of every moment that requires judgment, taste, or signature. The platform handles every step in between.
Start a pilot
Free for 14 days. Invite-only. No credit card required.