We have been waiting for ourselves there for some time, and now it’s here: Openai has introduced an agent coding tool called Codex in Research Overview. The tool is intended to allow experienced developers to delegate programming tasks by heart and relatively simple to an AI agent who will generate the code ready for production and show his work along the way.
Codex is a single interface (not to be confused with the CAD Code tool introduced by OPENAI last month) which can be reached from the sidebar of the Chatgpt web application. Users enter an prompt, then click on “Code” so that it begins to produce code, or “asks” to make him answer questions and advise.
Whenever a task is given, this task is performed in a separate container which is preloaded with the user’s code base and is intended to reflect their development environment precisely.
To make the codex more efficient, developers may include an “agents.md” file in the deposit with personalized instructions, for example to contextualize and explain the code base or communicate standardization and style practices for the project – as Readme.md but for AI agents rather than humans.
The Codex is built on Codex-1, a refined variation of the O3 reasoning model of OpenAI which has been formed using learning by strengthening on a wide range of coding tasks to analyze and generate code, and to iterate tests along the way.
