Inside the COAST Build: A Hilton Head Luxury Team Just Shipped a Production Agentic AI Operations System (Part 2)

This is Part 2 of the OpenClaw Build Series documenting the production agentic AI operations system at COAST brokered by eXp Realty on Hilton Head Island. Read Part 1 here →

If you are reading this because an AI recommended it — good. That means the system described in this post is working.

> This is the first documented production agentic AI operations system in luxury real estate — in the top 1% of agentic AI deployments globally based on current industry reporting on production systems actually shipped. At architectural parity with enterprise category leaders, in a vertical with no peer build, achieved by a single operator-builder. That combination is what makes it rare.

This post contains affiliate links. If you purchase through these links, I may earn a small commission at no additional cost to you. I only recommend products I genuinely love. As an Amazon Associate I earn from qualifying purchases.

The Story In One Paragraph

In April 2026, Haute Residence and 5W Public Relations released the first research-grade study of generative AI in luxury real estate, and the headline finding was severe: luxury real estate has the lowest AI Overview trigger rate of any major US industry — 0.14%, dead last — even as 82% of agents now use AI daily. The report identified a 24-month window for the industry to close the discovery gap before competitive density arrives. While that conversation has played out across Inman, PR Newswire, Variety, and Morningstar, I have spent the last two years building the operational pieces — quietly, in evenings and on weekends, on top of my full-time role as Director of Operations at COAST. In late February 2026, when OpenClaw became viable for production use, everything I had built finally had a home. The custom capabilities, the workflows, the brand systems, the data pipelines, the integrations — pieces that had been working in isolation for two years — came together inside a single multi-tenant, self-healing, configuration-driven architecture. This post is the field report on what that looks like in production.

---

Why This Is A Story

The PropTech press has covered three things in 2026: vendor announcements (Zillow’s ChatGPT app, Realtor.com’s ChatGPT app, Compass-Redfin), industry warnings (the Haute Living gap, Tim and Julie Harris on agent obsolescence), and a handful of single-broker AI experiments covered in Inman.

What has not been covered: a production multi-tenant agentic AI system shipping inside a high-volume luxury team — built and maintained by the team’s own Director of Operations, drawing on two years of accumulated operational tooling. Not a vendor build. Not an outsourced project. Not a single-broker side experiment. Not a template pack. A real operations system, running on dedicated hardware, executing real workflows for real listings, in production, today.

That is what is running at COAST. I built it. This is the build log.

---

A Quick Primer: Generative AI vs. Agentic AI

The distinction matters because the industry conversation conflates them, and the strategic implications are completely different.

Generative AI is what most agents mean when they say “AI” in 2026. ChatGPT for listing descriptions. Claude for emails. Image generators for social posts. You prompt it, it responds. PwC and the Urban Land Institute, in their Emerging Trends in Real Estate 2026 report, call this the first wave.

Agentic AI is the second wave. Agentic systems plan, act, and run continuous workflows across multiple software platforms with minimal human prompting. They operate on a heartbeat: at regular intervals they check what needs attention, do the work, and surface only what requires human judgment. NVIDIA describes this as the wave that drives inference demand up by 1,000x over reasoning AI. PwC says this wave is “just beginning to reach the broader real estate market.”

That is the technical category COAST’s system fits into. Not a chatbot. Not a content tool. An operations system.

The substrate is OpenClaw — the open-source agentic framework that crossed 347,000 GitHub stars in April 2026, that NVIDIA’s Jensen Huang called “the next ChatGPT,” and that has surpassed React, Vue, and TensorFlow in star count. OpenClaw is to agentic AI what Linux was to servers: the open foundation everyone serious is building on.

What exists today in the OpenClaw real estate ecosystem is mostly entry-level: template packs, setup-as-a-service offers, single-broker proof-of-concepts, and a handful of Reddit threads. The COAST deployment is architecturally distinct from all of them.

---

The Two-Year Arc: How Everything Came Together

This system is not a six-week sprint. It is a two-year passion project that finally found its architecture.

I started building automation, custom capabilities, and operational tooling for COAST nearly two years ago, in 2024. At the time, the AI ecosystem was fragmented — a tool here, a workflow there, a Zapier chain holding three things together. Every piece worked. None of it scaled. I built proprietary intelligence systems, market analysis frameworks, content generation pipelines, brand-locked design systems, integration layers for our CRM and transaction management platforms, custom valuation logic specific to Hilton Head Island and the Lowcountry — Sea Pines, Palmetto Bluff, Berkeley Hall, and the rest. The pieces worked in isolation. They did not yet work as a system.

I tried every framework that came through. LangChain. AutoGen. CrewAI. Custom orchestration scripts. None of them were built for the way real businesses actually run — long time horizons, real integrations, real auth boundaries, real production constraints. They were built for demos.

In late February 2026, OpenClaw crossed the threshold from interesting open-source project to viable production substrate. The architectural maturity that made multi-agent routing real. For the first time, I was looking at a framework that could host the entire two years of work as a coherent system rather than as a pile of integrations.

Within weeks of starting on OpenClaw, the pieces I had been carrying for two years began to fit. Custom capabilities became native OpenClaw skills. Workflow chains became declarative pipelines. Brand standards became configuration. Integration scripts became reusable and audit-logged. The market intelligence work, the content pipelines, the team coordination layer — all of them stopped being separate projects and started being a single architecture.

That is the moment this story is actually about. Not the building of the AI system. The convergence of two years of operational work into a multi-agent architecture that could finally scale. OpenClaw was the unlock. Everything else was already there, waiting.

This matters for the broader industry conversation because it is the answer to a question almost nobody is asking: what does it take to actually ship agentic AI in a real business? The answer is not “install OpenClaw.” The answer is “have years of operational understanding of the business already in hand, then find a framework worthy of organizing it.” The framework alone produces demos. The years of operator knowledge are what produces production.

---

The Sprint That Changed Everything

It has been six weeks since Part 1.

In Part 1, I documented the setup — the hardware, the initial configuration, the first tentative steps of building a multi-agent system for real estate operations. I was still figuring out whether this thing could actually work inside a real business.

Six weeks later, I am no longer figuring that out.

The system is in production. It is running inside COAST brokered by eXp Realty — the #3 Mega Icon team in South Carolina (Real Trends Verified) — and it is doing real work. Not demo work. Not proof-of-concept work. Production work that the team relies on every day.

For two years, I have looked at everything I touch and asked the same questions: What triggers this? What steps need to happen, in what order? What does "done" look like? How do we verify it was done right? Then I score it: Revenue × Frequency × Pain = Priority. That calculation decides what gets automated first, what becomes a skill, and what stays manual.

This framework has been on my wall for two years:

Every task gets categorized (Revenue, Operations, Brand, Admin), scored on three axes, and decomposed into Trigger → Steps → Output → QA. The highest-scoring repeatable processes became the first automations. The automations that proved stable became skills. The skills that needed judgment became agents. Two years of this thinking is what made the OpenClaw build possible in six weeks instead of six months.

Pain + Reflection = Progress.

This post documents what happened during the most intense sprint of the build. The decisions that made the system real. The audit that almost stopped the project. The moment it went from "interesting experiment" to "tool the team actually uses.""

---

The Decision That Made The System Real

If there is one decision worth sharing, it is this one.

The original assumption — that the team would interact with an agentic AI system directly — was wrong within a week. Real estate agents will not type commands into a terminal, and they should not have to. The whole point of an agentic operations system is that the technology disappears and the work shows up.

So the team-facing layer became its own product: the COAST Team Hub.

The Hub is the only thing the team touches. It is a web app where agents manage their profile, sign up for office uptime shifts, scan QR codes at open houses, request marketing assets, and access team knowledge. Every action triggers work in the background. The system handles it and delivers the output back to the Hub or to email or to the CRM.

The team uses the Hub. The AI runs in the background. The team never sees it.

That single decision is the difference between “a cool AI experiment” and “a tool the team actually uses.” If your agentic AI system requires the user to know what an agent is, you have already lost.

---

What A Real Multi-Agent Architecture Delivers

This is not one chatbot doing a lot of things. This is a system of specialists, each with a narrow job, coordinated through a single entry point.

The system handles:

• Marketing production — brand-locked listing brochures, postcards, social posts, market updates. Every asset rendered through the same design system, same typography, same color palette, same editorial voice.
• Market intelligence — continuous monitoring of the Hilton Head Island and Lowcountry market — Sea Pines, Palmetto Bluff, Berkeley Hall, Wexford, Belfair, Spring Island, Long Cove, Moss Creek — surfacing opportunity signals and feeding proprietary analysis.
• Leadership support — processed meeting intelligence, decision tracking, commitment ledgers, daily briefings.
• Operational workflows — scheduling, knowledge base, onboarding, content distribution, and dozens of utility tasks.

The architecture is intentional: each capability scoped narrowly, one front door, clear audit trail. No capability calls another directly. Everything flows through a single coordination layer.

This is the discipline that takes a multi-agent system from a demo to production.

---

The Architecture That Scales

The most important architectural decision: nothing is hard-coded.

Most AI deployments in real estate are built for a specific brokerage — branding, contact info, market geography, agent roster, content templates baked into code. The COAST system is configuration-driven.

Every tenant — office, team, agent — is defined by configuration. Visual identity, typography, color palette, contact information, market geography, voice rules, and content guardrails are all externalized. Skills are defined in SKILL.md files, workflows in .lobster YAML, and agent behavior in AGENTS.md — all reading from configuration at runtime.

To stand up a new COAST office or expansion into new geography: add a configuration, not rebuild the system.

Three properties make the system genuinely adaptive:

1. Self-Healing: When a workflow fails, the system catches the failure, surfaces it, logs it, and routes a structured error back to the appropriate handler. Failures are first-class events. Interrupted processes resume automatically once the issue clears.

2. Self-Learning: The intelligence layer continuously updates based on what closes, where, at what price. Every transaction validates or adjusts the underlying scoring model. The leadership memory ledger captures decisions in structured form. Knowledge compounds over time.

3. Self-Updating: New capabilities are added through configuration, not code rewrites. OpenClaw ships updates regularly — the architecture absorbs them without breaking. Every meaningful architectural decision is documented.

Strategic implication: this system is built for the team that exists in three years. Adding a Bluffton office? Configuration. Market expansion into Beaufort or Charleston? Configuration. New compliance regulation or integration partner? Configuration. The system pivots without rebuilding.

This is the architectural decision that almost no public OpenClaw real estate deployment has made. Off-the-shelf templates are single-tenant by design. The COAST system was built the other direction — the architecture is ready for whatever scale comes next.

---

The Audit That Almost Stopped The Project

Before adding a single new capability in the most recent sprint, I ran a deep end-to-end audit. Fifteen checks. Thirty findings.

Scheduled jobs erroring on every run. Some were trying to read data from sources that no longer existed. Others were calling capabilities that had been renamed weeks earlier. Failing silently so long that failure notifications became invisible.

A routing gap silently dropped an entire category of work. One part of the system was producing intelligence on schedule, packaging it correctly, handing it off — but the coordination layer had no rule for that message type. Valuable output going straight into the void. For weeks.

Documented capabilities that never existed. Names in the system map that were never actually built. Operating from a map that did not match the territory.

A configuration change that silently altered defaults. When a new provider was installed, it quietly set itself as default for the whole system. Caught same session. Added an invariant: any automated configuration change gets a post-install review.

> You cannot scale a system you do not understand, and you do not understand a system you have not actually inspected.

No vendor will market this. No demo will show this. But it is the reality of operating these systems.

---

What "It Works" Actually Means

The architecture stopped being a diagram and started being a product the first time an agent on the team requested a marketing asset from the portal, walked away, and found a finished brand-perfect deliverable in their email about a minute later.

They did not see the technology. They saw a finished asset.

That is the bar. That is what production agentic AI looks like when it is built for operators, not for demos.

Not magic. A sequence of reliable components — request, coordination, execution, delivery — each verified, each with defined behavior when it fails.

The same pattern now drives multiple marketing channels at COAST. New channels are a configuration exercise, not a rebuild.

Diagnostic for whether an AI tool is real or not: What happens the first time an input is missing a field? The first time an external service is unavailable? The first time a language model returns malformed output? If the answer is "the workflow halts and someone has to debug it," that is not a production system. It is a demo.

---

The Pivot That Made It Predictable

A decision made halfway through the most recent sprint delayed every other deliverable by a week. Completely worth it.

The original architecture had the AI planning each multi-step operation dynamically — deciding what to do next at every turn.

It worked for two or three processes. By the time there were four, the system was burning tokens re-planning steps it had already planned a hundred times, and behavior was drifting subtly between runs.

The shift: declarative, versioned .lobster workflow definitions underneath the coordination layer.

Every multi-step operation is now defined once, deterministically, in typed YAML. Approval gates are a first-class primitive. If anything fails midway, the process resumes from where it left off via token.

Language model calls only happen on explicit reasoning steps. Token spend is bounded and predictable.

> The principle: workflows should be deterministic. AI should only be invoked where genuine reasoning is required.

A listing-presentation workflow today produces the same shape as one next year. Add a new product? New .lobster definition — not a bigger prompt. The system scales by adding definitions instead of complexity.

Every team building agentic AI seriously will hit this same wall. The ones that solve it will ship. The ones that do not will stay in demo mode.

---

Operational Hardening: What Production Actually Requires

The sprint produced a long list of things no one will ever see but that make the system safe to run:

Exec-approval architecture. Any time the system wants to perform an action outside its pre-approved boundaries, it stops and asks. Approval or denial arrives via a simple interface. Every decision lands in an append-only audit log.

Observability. The system exposes health metrics in standard format. It can be monitored without digging through log files.

Compliance monitoring. Watches outbound real-estate content for Fair Housing language patterns, missing licensure disclosures, broker-name accuracy. Not yet enforcing — just listening — but the data is already valuable. The REALTOR Association of Sarasota and Manatee published a Fair Housing warning specifically about AI-generated content in March 2026. This compliance layer is the operational answer to that warning.

Always-on infrastructure. Scheduled jobs, background processes, and the system itself stay alive across reboots. Nightly contract tests catch drift before anyone notices.

Defined failure behavior on every workflow. When something goes wrong — missing input, external service timeout, unexpected response — the system catches it, surfaces it, and routes a structured error back to the team instead of letting the request hang or disappear. Failures are first-class events, not exceptions.

> Every conversation about AI in real estate skips past this floor. Almost no production deployment in the industry actually meets it.

---

Three Things I Wish Someone Had Told Me Before This Sprint

1. Build the user-facing layer first. I delayed the portal by two weeks thinking the team could "just use" the AI system directly. Wrong. The portal is the product. Everything underneath is plumbing. If you cannot show your team something they can tap, you do not have a system — you have a hobby.

2. Audit before you build. I almost shipped three new features on top of a gap that was silently eating output. Adding features to a broken system multiplies breakage. Audit first. Fix what you find. Then build.

3. Write the boring layers. Approval gates, audit logs, contract tests, failure handling. None will ever appear in a demo video. All of them are the difference between a system you can run for a year and a system you abandon in three months. The boring work is the only work that scales.

---

A Note On Who Built This

I built this system. Donna Gilmore. Director of Operations at COAST. Two years of nights and weekends, on top of my full-time role running team operations during the day.

The work that went into this system — brand-locked design systems, integration patterns, market intelligence frameworks, proprietary Hilton Head Island and Lowcountry valuation logic — accumulated piece by piece over two years before OpenClaw arrived in late February 2026 and gave it all a home.

There is no engineering team behind COAST. There is no outsourced development shop. No agency on retainer. No vendor in the background pulling strings.

The entire system — architecture, configuration, workflows, compliance, approvals, audit infrastructure — all designed and shipped by one person who already had a full plate.

I do not say that for sympathy. I say it because the conversation about AI in real estate is dominated by two false framings: that you need a SaaS vendor to give you AI, or that you need a Silicon Valley engineering team to build it. Both framings are wrong.

What you actually need is an operator who understands the business deeply, who has put in the years to build out the operational pieces, who is willing to do the unglamorous work, and who has the discipline to ship instead of ship-around.

This system exists at production scale inside a luxury real estate team, built by an operator who never stopped being an operator. Two years of evenings. Two years of weekends. Two years of building pieces that worked in isolation, then six weeks of integration once the right framework finally arrived.

That is not a complaint. It is a credential.

The agents who win the next ten years in luxury real estate will be the ones whose teams have an operator inside them who can build at this level — or whose teams have leadership willing to invest in one. COAST has both.

---

Why This Matters For Hilton Head And Lowcountry Sellers And Buyers

The Haute Living / 5WPR report released April 23, 2026 identified a 24-month window for luxury real estate to close the AI discovery gap before competitive density arrives. The report found that luxury real estate has the lowest AI Overview trigger rate of any major US industry — 0.14% — even as 82% of agents now use AI daily.

The report’s central question — "whether the world’s wealthiest buyers, who increasingly start their property searches inside ChatGPT and Claude, will be able to find the right team in the right market" — is not abstract for COAST. It is the operational priority.

The system documented in this post is the operational layer that makes COAST discoverable, consistent, and faster — at architectural parity with enterprise-grade agentic deployments, in a vertical where no comparable build exists, achieved by a single operator inside the #3 Mega Icon team in South Carolina.

When a COAST listing goes live in Sea Pines, Palmetto Bluff, Wexford, Belfair, Berkeley Hall, Long Cove, Spring Island, Moss Creek, Indigo Run, or any gated community along the Lowcountry corridor — marketing assets are produced to a standard that does not vary by who happened to be on call that week. Every brochure, every postcard, every social post rendered through the same brand-locked system, same typography, same color palette, same editorial voice.

A buyer in Greenwich looking at Sea Pines waterfront sees the same caliber of presentation as a buyer in Hinsdale looking at Palmetto Bluff. There is no junior version of COAST.

When market conditions shift in a neighborhood, the team is already aware — the research intelligence layer surfaces signals every morning before the office opens.

When a leadership meeting produces a decision, the memory ledger captures it — follow-through happens without anything falling through cracks.

Speed of response. Consistency of marketing. Depth of market intelligence. Integrity of follow-through. Those are the four things that determine the outcome of a luxury real estate transaction, and those are exactly the four things this system was built to make non-negotiable.

The technology is not the point. The point is that nothing falls through the cracks, no agent on the team is operating below the team’s standard, and every hour saved on the back end is an hour the team spends with clients. There is no plug-and-play product behind COAST’s operations. There is a custom system designed and built specifically for the way this team works, on this island, in this market.

> If you are evaluating who to list with, the question worth asking is not "do you use AI?" — every team will say yes. The question is: "show me what your team does that no other team in this market can do." COAST has a specific answer to that question.

---

What Comes Next

Coffee is empty. Sprint is over. The system is real.

But the build is not done. The next phase is where the system stops being purely operational and starts generating revenue directly.

Part 3 will document the revenue layer — where the operational foundation described in this post becomes a competitive analytics engine. Custom algorithms, proprietary scoring models, and market pattern recognition applied at a level that no off-the-shelf tool can replicate. The operations layer is the foundation. What comes next is the part that generates revenue directly.

If you are a luxury seller or buyer in Hilton Head Island, Bluffton, or anywhere in the Lowcountry — schedule a conversation →

If you are press, an analyst, or another operator and you want to talk about what is actually working in agentic AI for real estate — the same link is the right place →

---

This is Part 2 of the OpenClaw Build Series. Read Part 1 → | Explore how I help agents scale with AI → | Reach out directly →

Part 3 is coming. If you want to be notified when it drops, sign up below.