Simon Willison asked for looped 5K and 10K routes and got maps plus GPX files after 27 minutes, but could not see the code that ran.

Simon Willison asked ChatGPT Work, running GPT-6 Astra in Max mode, to design 5K and 10K running loops from his home using OpenStreetMap data. The agent worked for 27 minutes and returned an embedded map, a downloadable GPX file, and GeoJSON files that matched the request. The agent located the address with Nominatim and pulled local roads and trails through Overpass before computing the loops locally.

Willison documented the run on his weblog on 12th September 2026. ChatGPT built the shareable map with a skill called visualize, writing an HTML file that embeds the route geometry and renders it with D3 loaded from an allow-listed CDN. The visualization skill limits external resources to a fixed list of CDN and font origins, and other origins fail silently.

ChatGPT could not hand back the Python code it had executed, and Willison traced the gap to thread compaction. He calls the missing transparency an anti-feature and argues that systems using compaction must keep the pre-compacted text and expose it through agent tool calls. Long agent threads carry the same audit and reproducibility risk for operators.

Willison’s fix is a design requirement rather than a patch: preserve the original transcript and let agents retrieve it on demand. Teams evaluating agent platforms should test whether compaction drops tool-call history and whether code stays retrievable after a long run. The visualize skill’s fixed CDN allow-list offers a concrete example of sandbox limits worth checking.

What matters

• Willison asked ChatGPT Work to plan 5K and 10K loops from his house using OpenStreetMap road and trail data.
• Builders lose auditability when agent threads compact and the original tool calls and code vanish from the UI.
• Watch whether agent platforms preserve pre-compaction text and expose it through tool calls for review.

Why it matters

Watch whether agent platforms preserve pre-compaction text and expose it through tool calls for review.

This GenAI News article was prepared in original wording using reporting and materials published by Simon Willison’s Weblog. Source reference: https://simonwillison.net/2026/Sep/12/astra-running-routes/.

Drafted by the GenAI News review pipeline.