Summary
OpenAI engineers Frill and Vivi show what happens when coding agents are given a goal and left to work for days. Vivi demonstrates Codex running with the Astra model on a remote machine to build an explorable 3D Ferrari F1 car from just seven reference photos, and a fly-through of the 21-kilometer Nürburgring Nordschleife with all 70 turns, which took a little over two days, mostly spent waiting for Blender renders. Codex used sub-agents for research and planning, and resumed from saved frames after RAM and overheating interruptions.
Frill contrasts this with the heavily prescriptive “exec plans” he presented last year, when a seven-hour run felt remarkable. Today he gives the model a goal, the trade-offs he cares about and a definition of good, and lets it work out the steps. His main example is a CI-maintenance agent, called Frillbot, that has worked for more than 1,000 hours, reproducing intermittent test failures, collaborating with researchers on a suspected faulty GPU, and spawning reviewer agents before asking a human to approve a fix.
The session closes with six Codex practices for long-running work, including one thread per outcome and the slash goal command, giving developers a practical template for setting up durable agents.
📺 Source: OpenAI · Published November 11, 2026
🏷️ Format: Workflow Case Study
