Michael Bolin outlines the core processes behind the Codex CLI, a software agent designed to make reliable software changes efficiently. Central to this system is the "agent loop," which orchestrates how the user interacts with the model and the tools it uses. Initially, the agent takes user input and formulates it into a prompt for the model. This prompt is then processed through a series of inference steps that involve converting text into tokens, which the model uses to generate a response. The agent loop continues, allowing the model to request tool calls for further actions, such as executing commands, until it produces a final response for the user. The article details how the Codex CLI communicates with the Responses API for model inference. Depending on the type of authentication—whether through ChatGPT or API-key—the CLI uses different endpoints to access the model's capabilities. When a user inputs a query, the Responses API organizes it into a structured prompt based on predefined roles, such as 'system,' 'developer,' 'user,' and 'assistant.' Each role has a specific weight that influences how the model interprets the input, ultimately guiding it toward generating a coherent output. Bolin emphasizes the iterative nature of the agent loop. The model's responses can lead to additional tool calls, allowing it to gather necessary information from the local environment. Each cycle ends with an assistant message that indicates the completion of a task, marking a clear handoff back to the user. This process reflects a sophisticated interplay between user commands and the model's capabilities, showcasing how Codex efficiently adapts to user needs while executing complex tasks in real-time.