Where should an independent coffee shop open in Houston?

880 H3 hexes (about 5 km² each) across Harris County, scored two ways: an expert-weighted overlay of census demographics and competition, and Esri's USA Geodemographic Embeddings ("places like where independent shops already are"). Both were tested on shops hidden from them. A manager agent and five subagents answer questions through a local MCP server; five saved runs are below.

How well does each method find held-out shops?

Five spatial blocks (k-means on hex centroids). For each block, its shops were hidden, the embedding query was rebuilt from the other four blocks and the overlay's competition term was recomputed without them. AUC compares shop hexes with non-shop hexes in the block (0.5 = random); P@k is the share of shop hexes among the block's top k, where k is the number of shop hexes. Mean, with the range across the five blocks in brackets.

Five saved agent runs

Recorded live on the run date shown; the page itself makes no API calls. Each run lists the delegations and MCP tool calls, then the summary agent's answer. Run 5 is a question the data cannot answer.