Run long horizon tasks with Codex | OpenAI Developers

This post details an experiment stress-testing the long-horizon capabilities of GPT-5.3-Codex by tasking it with building a design tool from scratch over 25 hours. It highlights the importance of durable project memory through structured markdown files (spec, plan, implement, documentation) for maintaining coherence and enabling repair. The post emphasizes the agent loop (plan, edit, run tools, observe, repair, update docs) and its reliance on real feedback, externalized state, and steerability. It also introduces the Codex app's features like parallel threads, skills, automations, and Git worktrees as crucial infrastructure for managing long-running tasks. The core contribution is demonstrating and analyzing the techniques that enable AI agents to reliably complete complex, multi-step coding tasks end-to-end with minimal human intervention.