Notably, both OpenAI and Anthropic publish their coding CLI clients as open-source projects on GitHub, allowing developers to inspect the implementations directly; they do not, however, make the ChatGPT or Claude web interfaces available in the same way.
An official peek inside the loop
Bolin’s post concentrates on what he dubs the “agent loop,” the central logic that coordinates interactions among the user, the AI model, and the software tools the model uses to carry out coding tasks.
As we wrote in December, every AI agent revolves around a recurring cycle. The agent receives the user’s input and crafts a textual prompt for the model. The model replies, either delivering a final result for the user or asking to invoke a tool (for example, running a shell command or reading a file). When the model requests a tool, the agent runs that tool, adds the resulting output back into the prompt, and sends the updated prompt to the model again. This loop continues until the model ceases to request tools and instead returns an assistant message for the user.
The looping sequence needs an origin, and Bolin’s post shows how Codex assembles the initial prompt passed to OpenAI’s Responses API, which performs model inference. The prompt is composed of multiple elements, each tagged with a role that sets its precedence: system, developer, user, or assistant.
The instructions field is sourced either from a configuration file provided by the user or from default instructions packaged with the CLI. The tools field specifies the functions the model may invoke — such as shell commands, planning utilities, web search, and bespoke tools exposed via Model Context Protocol (MCP) servers. The input field holds a collection of entries detailing sandbox permissions, any optional developer directives, environment context (for example the current working directory), and ultimately the user’s message.
