An AI Flagged Four Expert-Selected Sites Sitting in Flood Zones. Nobody Else Caught It. A new paper on arXiv describes AURA: a multi-agent AI system that selects affordable housing sites under 127 regulatory constraints. Not filters applied after the fact. Constraints baked into the optimization. Zoning. QCT eligibility. LIHTC compliance. Flood zones. All of it. The architecture is what matters here. Four specialized agents: one reads geospatial data through graph neural networks. One checks every regulatory box in real time. One optimizes across four competing objectives (transit access, environmental impact, cost, social equity). And one coordinates the tradeoffs when agents disagree. Regulation always wins the tiebreaker. Results across 8 U.S. cities and 47,000+ parcels: 37% better outcomes across all four objectives compared to human expert selection. 94% regulatory compliance. Site selection compressed from 18 months to 72 hours. Selected sites scored 31% better on transit access and 19% lower on environmental impact than what the experts picked. In the NYC deployment, AURA found 23% more viable sites than planners did. Three locations in Astoria and Sunset Park were initially overlooked but scored Walk Scores of 88-91 at half the land cost of Manhattan. It also flagged four expert picks sitting in 100-year flood zones. Here’s what is interesting to me. This is specialized agents collaborating on a real allocation problem. The kind loaded with constraints. That’s not just affordable housing. That’s multifamily land acquisition. Student housing siting. Data centers. Any development where you’re juggling zoning, cost, access, and risk at the same time. Tools like Claude Code now make this kind of multi-agent coordination possible inside your own workflows. The interesting shift isn’t better models. It’s orchestration. Specialized agents that debate, challenge, and reach answers no single analyst would find. Real estate runs on constraints. AI is built for them. We’re getting closer to tools that don’t just summarize deals. They design them. I know what I’m building this weekend in Claude Code. What about you? Paper: arxiv.org/abs/2602.03940