SCALING LONG-RUNNING AUTONOMOUS CODING

Summary from the TLDR newsletter

Model choice matters for extremely long-running tasks. GPT-5.2 models perform better at extended autonomous work compared to Opus 4.5 and GPT-5.1-codex. Many improvements come from removing complexity rather than adding it. The best system is often simpler than you'd expect. The harness and models matter, but the prompts matter more.