How OpenAI Codex Works

OpenAI Codex is a cloud-based coding agent built around codex-1, a fine-tuned version of o3. The real engineering challenge was the orchestration layer, not the model itself. Three key layers define the system: the agent loop (iterative inference and tool execution), prompt and context management (layered prompt construction, quadratic growth mitigation via prefix caching, and context compaction), and a multi-surface architecture. When MCP proved insufficient for rich agent interactions like streaming progress and mid-task approval pauses, OpenAI built a custom JSON-RPC App Server protocol that powers the CLI, VS Code extension, web app, desktop app, and third-party IDE integrations from a single codebase. Practical takeaways include writing AGENTS.md files for project context, scoping tasks tightly, and starting fresh threads to avoid context degradation.