OpenAI Details the Codex Agent Loop for Local AI Coding

OpenAI has published a detailed explanation of the Codex agent loop, as outlined on its website. The post describes how the Codex CLI manages the cycle of user input, model reasoning, and tool execution to create and modify software on a local machine.

The agent loop is the core logic that connects the user, the model, and the tools the model calls to complete coding tasks. Each cycle, known as a turn, begins with user input that forms part of a prompt sent to a model through the Responses API. The model then generates a response, which can be either a direct reply or a tool call. When a tool call is made, the agent executes it, appends the result to the prompt, and re-queries the model until a final assistant message is produced.

Codex CLI supports multiple configurations for running inference, including OpenAI-hosted endpoints, ChatGPT’s backend API, and local setups through OSS models such as Ollama or LM Studio. It constructs prompts by combining system, developer, and user messages, and includes environment context such as working directory and shell information.

The post also details how Codex handles performance and context management. To maintain efficiency, prompt caching is used to avoid redundant computations, while a compaction process automatically summarizes long conversations when token limits are reached. These mechanisms help Codex operate efficiently without compromising its stateless and zero-data-retention design.

This explanation follows the company’s earlier introduction of the Codex CLI, which serves as a local agent for developers to automate coding workflows using large language models.