Uncovering How the Codex Agent Loop Works

Codex CLI is a software agent designed to make software changes safely and efficiently in the user's local environment. At the heart of this system is the agent loop, which coordinates the interaction between the user, the model, and the tools the model calls.

The agent loop works through the following process:

• Prompt generation: prepares a set of instructions that includes the user's input.
• Inference: passes the instructions to the model to generate a response.
• Tool Call: when the model requests execution of a specific tool (e.g., running the ls command), the agent performs it and adds the result back into the prompt.
• Iteration and termination: this process repeats until the model sends a message directly to the user (Assistant message) instead of a tool call.

The process from the user's input to the agent's final response is called a turn. Within a single turn, multiple rounds of inference and tool calls can occur, and as the conversation progresses, the prompt grows longer, making management of the model's context window an important design consideration.