For most of us an “AI agent” is still a black box. This 3-hour workshop opens it. We train no model: we hand an existing LLM the tools of GIS — SQL, maps, charts — and assemble an “AI as analyst”: the conversational, tool-using corner of GeoAI.

The vehicle is geo-chat, an open-source app built for this workshop, running entirely in the browser with no server: DuckDB-WASM (spatial extension) for SQL, MapLibre GL rendering vector tiles generated on the fly via ST_AsMVT, Vega-Lite for charts, and Anthropic Claude operating all of it through tool calls.

The method is “take it apart to understand it.” After seeing the finished app, we switch to a branch with every capability subtracted, then add the layers back one at a time — tools, visualization, validation, skills, evals — throwing the same request at each: all talk; SQL that runs but paints nothing; a map that paints and silently breaks; a break that becomes a readable error; a skill that fixes a real coordinate-order bug. Same task, same model, same key: every difference comes from one layer.

Along the way we watch the agent loop in the browser’s Network tab as raw API traffic — including the surprise that the model is stateless, the whole conversation re-sent every call. We see why declarative specs suit AI, and that a skill is one Markdown file loaded only when needed. The finale is evals: participants run the real loop against the real API and write one automated evaluation, because with a non-deterministic agent the evaluation is the product.

No implementation expertise is required. You leave able to design tools and skills for your own data, and to debug an agent at the API-request level. Materials are open source, English and Japanese.