tech
January 27, 2026
OpenAI spills technical details about how its AI coding agent works
Unusually detailed post explains how OpenAI handles the Codex agent loop.
TL;DR
- Michael Bolin, an OpenAI engineer, published a technical breakdown of the Codex coding agent.
- The "agent loop" orchestrates user interactions, AI models, and tools for coding tasks.
- The system constructs prompts from system, developer, user, and assistant components.
- Codex handles prompt growth and performance issues with techniques like prompt caching.
- Conversations are automatically compacted when token counts exceed a threshold.
- Both OpenAI and Anthropic open-source their coding CLI clients.
- AI coding agents are useful for prototypes but require human oversight for production work.
- Future posts will cover CLI architecture, tool implementation, and sandboxing.