Environmental Massing Agents — Design skill for Claude Code

Environmental Massing Agents

A multi-agent computational design **study**, not just a generator: five Claude agents (a Generator and four domain Critics — solar, embodied carbon, pedestrian wind, and view-corridor/zoning yield — plus an Orchestrator that synthesizes their critiques) negotiate a real, physically-grounded design space through a genuine **Model Context Protocol (MCP) server**, the same integration pattern documented in Xun Liu's own NeurIPS 2025 paper on MCP-driven parametric modeling for Grasshopper. On top of that system, this project states four formal research questions and falsifiable hypotheses **before** running anything, then answers every one of them with real, reproducible computation: a real evolutionary multi-objective optimizer (NSGA-II) benchmarked against random sampling with a real performance metric (hypervolume) over 5 random seeds each, a closed-form geometric prediction derived from convex geometry and checked point-by-point against simulation, an exhaustive 1,080-design controlled comparison, and a correlation analysis of the resulting Pareto-optimal set. One hypothesis was falsified in its strict form; that result, and why, is reported here rather than hidden, because a hypothesis that survives contact with real computation only when stated loosely is itself a finding.

**This is a Rhino/Grasshopper study with a live MCP server, not a web app.** See "How to test it" below for exactly what to run and what each command should print or produce.

Why this, as a fourth project

The first three projects in this series each demonstrated one capability: reading zoning into a form, answering zoning questions with citations and honest refusal, and computing a single hard geometric constraint correctly with a proven research finding. This project asks a harder question none of the first three could, and, unlike a project that stops at "here is a system that runs," it commits to specific, falsifiable predictions about that system's own