Generating running routes from OpenStreetMap data with GPT-6 Astra

Summary

Simon Willison used ChatGPT Work (GPT-6 Astra Max) to generate 5K and 10K running routes starting from a home address, leveraging OpenStreetMap data. The process took 27 minutes; the agent produced GPX and GeoJSON files alongside an embedded D3 visualization. Willison highlights a transparency issue: the code generated by the agent was not visible in the interface and became impossible to recover after the chat flow was compressed.

Why it matters

This is a concrete early example of agentic large language model (LLM) systems combining geographic APIs like Nominatim and Overpass with local calculations to deliver practical, multi-format outputs without requiring human coding. It also exposes a genuine usability gap in current agent products. When the system context is “compressed,” users lose access to the code and reasoning that produced the results. This undermines trust and reproducibility.

Highlights

• The agent provided GPX, GeoJSON, and an inline D3 map in a single pass, using Nominatim, Overpass, and local loop calculations.
• Flow compression caused the generated Python code to become permanently inaccessible to the user.
• The visualization relied on a “visualization skill” with a strict CSP allowlist for CDNs like jsdelivr and unpkg.

Critical take

Willison’s concern is justified. However, the core issue is architectural. A product that hides its own execution outputs and then deletes them through compression presents a black box mimicking a tool. This is a serious anti-feature for any user who needs to audit, reuse, or debug the output.