OpenAI has released a research preview of its agentic coding tool, Codex. We’ve waited a long time for this. The tool allows experienced developers to delegate repetitive and relatively simple programming tasks. An AI agent will generate production-ready codes and show its progress along the way.
Codex is an interface that can be accessed from the sidebar of the ChatGPT app. Users enter a prompt, and then click either “code” It will begin to produce code or “ask” It can answer questions and give advice.
It will perform any task that it is given in a container that has been preloaded with code from the user and is intended to accurately reflect that environment.
Developers can include an array of features to make Codex more effective. “AGENTS.md” File in the repo with customized instructions, such as to contextualize and describe the code base, or to communicate standardizations & style practices for the Project–kind of like a README.md for AI agents instead of humans.
Codex is based on codex-1. This is a finely tuned variation of OpenAI’s o3 reasoning models that was trained with reinforcement learning to analyze and generate code and to iterate throughout tests.
