OpenAI's Codex coding agent, while powered by a fine-tuned model, relies on extensive engineering beyond the AI itself. The team faced significant challenges in building the orchestration layer, particularly around handling large conversation histories and managing context windows. Codex uses an agent loop where user input triggers model inference, potentially followed by tool calls executed by the harness, which manages everything from file edits to test runs. This loop repeats, sometimes dozens of times, until a final user response is generated. The system's complexity extends to prompt management, where each user message becomes a layered prompt stacked with environment context, project-specific instructions, and tool definitions, leading to quadratic growth in data transfer. OpenAI accepts this cost to maintain statelessness and zero data retention, using prompt caching as a mitigation.

Compaction handles context window limits, replacing full history with encrypted state summaries. AGENTS.md files exemplify the design choice to externalize project-specific knowledge, allowing developers to define conventions within their repositories.