Generating Running Routes with ChatGPT Agents and OpenStreetMap Data

Simon Willison demonstrated using a ChatGPT agent powered by GPT-6 Astra to generate customized 5K and 10K running loop routes starting from a home address. The agent autonomously queried spatial APIs, wrote and executed route-calculation code, and returned interactive map visualizations alongside downloadable GPX and GeoJSON files. This showcases how modern LLM agents can execute complex, multi-step workflows that combine spatial data retrieval, custom algorithmic computing, and data export for real-world tasks. However, it also highlights usability issues around agent transparency and context loss when chat threads are automatically compacted. The agent used Nominatim to geocode the address and the Overpass API to download local road and trail networks from OpenStreetMap, calculating loop routes in a Python sandbox over 27 minutes. Willison noted a major limitation: because the ChatGPT UI hid the code execution details and compacted the thread, the generated Python script was lost and could not be inspected later.