It’s worth keeping in mind that both OpenAI and Anthropic open-source their coding CLI customers on GitHub, enabling designers to take a look at the execution straight, whereas they do not do the exact same for ChatGPT or the Claude web user interface.

A main appearance inside the loop

Bolin’s post concentrates on what he calls “the representative loop,” which is the core reasoning that manages interactions in between the user, the AI design, and the software application tools the design conjures up to carry out coding work.

As we composed in December, at the center of every AI representative is a duplicating cycle. The representative takes input from the user and prepares a textual timely for the design. The design then produces a reaction, which either produces a last response for the user or demands a tool call (such as running a shell command or checking out a file). If the design demands a tool call, the representative performs it, adds the output to the initial timely, and queries the design once again. This procedure repeats up until the design stops asking for tools and rather produces an assistant message for the user.

That looping procedure needs to begin someplace, and Bolin’s post exposes how Codex constructs the preliminary timely sent out to OpenAI’s Responses API, which manages design reasoning. The timely is constructed from a number of elements, each with an appointed function that identifies its top priority: system, designer, user, or assistant.

The guidelines field originates from either a user-specified setup file or base guidelines bundled with the CLI. The tools field specifies what operates the design can call, consisting of shell commands, preparation tools, web search abilities, and any custom-made tools supplied through Model Context Protocol (MCP) servers. The input field includes a series of products that explain the sandbox consents, optional designer guidelines, environment context like the present working directory site, and lastly the user’s real message.