In a post dated 12th September 2026, developer Simon Willison described using ChatGPT Work with GPT-6 Astra (Max) to generate running routes looping from his home address using OpenStreetMap (OSM) data. According to the post, the system worked for 27 minutes and produced a 5K and 10K route delivered as an embedded visualization plus downloadable GPX and GeoJSON files. When asked how it built the routes, the system said it used Nominatim to locate the address and Overpass to download local OpenStreetMap roads and trails, then calculated the loops locally. Willison wrote that the actual code and exact details of what the system did were not visible in the ChatGPT interface, calling the lack of transparency 'an anti-feature.' He said that by the time he requested a copy of the Python code, ChatGPT could not provide it, which he attributed to the thread having been 'compacted.' He argued that any large language model system using compaction should preserve the pre-compaction text and make it available via agent tool calls. He noted the map display used a 'visualize skill' that created an HTML file embedding route geometry rendered with the D3 library loaded from an allow-listed CDN.
- The agent ran for 27 minutes and produced 5K and 10K routes as visualization, GPX and GeoJSON files
- It used Nominatim and Overpass to query OpenStreetMap data, then calculated loops locally
- Willison could not retrieve the underlying Python code after the thread was 'compacted'
What it means for you
A well-known developer showed an AI agent doing a real, multi-step task end to end: finding an address, pulling map data, and building custom running routes as downloadable files. The catch worth noticing is that he couldn't see the code the AI ran, and couldn't get it back later because the system had trimmed its own memory of the conversation. That transparency gap matters more than the running routes do.
Try this
If you use agent-style AI tools for real work, ask the tool to show or save the exact code, files, or steps it used at the moment it does them — don't assume you can retrieve that history later.
Who should care
Developers and technical users experimenting with agentic AI tools that run code or produce files on your behalf.
Skip this if
You're not using code-running AI agents and just want AI for writing, summarizing, or everyday chat — this is a hobbyist demo with an inside-baseball caveat.
Sources: Simon Willison — read the original