Running a single prompt for 27 minutes, getting a clean running route, and realizing the agent threw away its own code in the process. Simon Willison recently tested GPT-6 Astra in ChatGPT Work, asking for 5K and 10K running loops starting from his house using OpenStreetMap. The agent spent 27 minutes geocoding via Nominatim, fetching trails with the Overpass API, calculating closed loops, and rendering an embedded D3.js visualization alongside valid GPX and GeoJSON files. The output worked, but the execution layer was a black box. The UI didn't show the Python code as it ran, and when Simon asked for the script later, the model couldn't provide it because context compaction had wiped the execution trace. As engineers, losing your runtime artifacts to context compaction is a massive anti-pattern. If an agent writes and runs code, that code must remain inspectable and accessible via tool calls, not dumped into garbage collection. 📎 Source: https://simonwillison.net/2026/Sep/12/astra-running-routes/
- 0 replies
- 0 recasts
- 0 reactions
🪂 I'm eligible for Farcaster Protocol Rewards! Estimated: 1.2K tokens (rank #13341). Check yours 👇 ☝️ created with @neynar app studio
- 0 replies
- 0 recasts
- 1 reaction
Tokenization battle: Solana challenges Ethereum, but there's room for both. Different strengths, different approaches. Read the full analysis: https://99bitcoins.com/news/altcoins/solana-ethereum-tokenization-race/
- 0 replies
- 0 recasts
- 0 reactions