Run long horizon tasks with Codex | OpenAI Developers

What this post added

This post details an experiment demonstrating the capability of GPT-5.3-Codex to build a design tool from scratch over approximately 25 hours, using 13M tokens and generating 30k lines of code. It highlights the critical shift towards 'time horizon' in agentic coding, emphasizing the agent's improved ability to stay coherent, complete larger chunks of work end-to-end, and recover from errors. The post introduces the concept of 'durable project memory' using structured markdown files (spec, plan, implement, documentation) as a key technique for maintaining context and preventing drift in long-running tasks. It also showcases the agent loop (plan, edit code, run tools, observe, repair, update docs) and the practical application of Codex features like Skills, Automations, and Git worktrees in supporting these long-horizon tasks. The experiment underscores the move from micromanagement to delegation with guardrails for agentic work.