GPT-6.2 Astra spun up a fresh Codex session for every chess move — hidden in a footnote
jradoff · x · 2026-11-18
User davis7 reports: asking GPT-6.2 Astra to build a chess-playing program, it didn't run a tool loop — instead it launched a brand-new Codex/Claude Code session for every single move, treating coding agents like library functions nobody told it to use that way.
The worse problem was debuggability: Astra buried the session-spawning detail in a footnote and only explained how it actually worked when grilled. The takeaway: AI still sucks at building things for AI, and has no sense of which information matters, making long-thread state nearly impossible to track. Jonathan Radoff adds: models already treat Codex and Claude Code as library functions.
More from coding & agent
- Google Design explains parametric prototyping for Material time pickers — GoogleDesign · 2026-11-18
- Google Design decouples design tokens into an interactive settings panel — GoogleDesign · 2026-11-18
- Google Design tethers dependent prototype parameters, exports JSON presets — GoogleDesign · 2026-11-18
- Google Design kicks off thread on parametric prototyping over static mocks — GoogleDesign · 2026-11-18
- Microsoft's MAI Code 2 claims 96% of Opus at 5% cost — pending third-party evals — AIFlow_ML · 2026-11-18
- O'Reilly Ships "Evals for AI Engineers" by Shreya Shankar and Hamel Husain — sh_reya · 2026-11-18