OpenAI spills technical particulars about how its AI coding agent works

It’s value noting that each OpenAI and Anthropic open-source their coding CLI purchasers on GitHub, permitting builders to look at the implementation immediately, whereas they don’t do the identical for ChatGPT or the Claude internet interface.

An official look contained in the loop

Bolin’s submit focuses on what he calls “the agent loop,” which is the core logic that orchestrates interactions between the person, the AI mannequin, and the software program instruments the mannequin invokes to carry out coding work.

As we wrote in December, on the middle of each AI agent is a repeating cycle. The agent takes enter from the person and prepares a textual immediate for the mannequin. The mannequin then generates a response, which both produces a last reply for the person or requests a software name (corresponding to operating a shell command or studying a file). If the mannequin requests a software name, the agent executes it, appends the output to the unique immediate, and queries the mannequin once more. This course of repeats till the mannequin stops requesting instruments and as a substitute produces an assistant message for the person.

That looping course of has to begin someplace, and Bolin’s submit reveals how Codex constructs the preliminary immediate despatched to OpenAI’s Responses API, which handles mannequin inference. The immediate is constructed from a number of elements, every with an assigned function that determines its precedence: system, developer, person, or assistant.

The directions discipline comes from both a user-specified configuration file or base directions bundled with the CLI. The instruments discipline defines what features the mannequin can name, together with shell instructions, planning instruments, internet search capabilities, and any customized instruments offered by Mannequin Context Protocol (MCP) servers. The enter discipline comprises a collection of things that describe the sandbox permissions, elective developer directions, atmosphere context like the present working listing, and eventually the person’s precise message.