We have been expecting it for a while, and now it is here: OpenAi has introduced an agentic coding tool called Codex in Research Preview. The tool is intended to enable experienced developers to delegate Rote and relatively simple programming tasks to an AI agent who will generate production-ready code and show his work along the way.
Codex is a unique interface (not to be confused with the Codex Cli tool that was introduced by OpenAI last month) that can be reached from the sidebar in the Web app Chatgpt. Users enter a prompt and then click on “Code” to let them start producing code, or “questions” to have questions answered and advised.
When it gets a task, that task is performed in a separate container that is in advance loaded with the user's codebase and is intended to accurately display their development environment.
To make Codex more effective, developers can include a “agents.MD” file in the repo with adapted instructions, for example to contextualize and explain the code basic or to communicate standardizations and style practices for the project -friendly of a readme.md but for AI agents instead of people.
Codex is built on Codex-1, a refined variation of the O3 reasoning model of OpenAi that was trained using reinforcement learning on a wide range of coding tasks to analyze and generate code, and by repeating tests along the way.
