OpenAI just revealed exactly how Codex works — and the model is the easy part

You know that feeling when you ask an AI to fix a bug and it gives you code that doesn't work because it can't run your tests? Before Codex: you'd paste code back and forth, manually run tests, and iterate. Now: Codex runs in an isolated sandbox with your full repo, executes tests, reads linter output, and keeps iterating until tests pass — all while you work on something else.

Think of Codex as a tireless junior developer in a sandbox. You give it a task ('fix the auth bug'), and it enters an agent loop: read files, form a plan, run commands, see what happens, adjust, repeat. Each turn can involve dozens of tool calls (shell commands, file edits, test runs) before it responds. The prompt stacks context like layers: system rules, your AGENTS.md instructions, sandbox permissions, tool definitions, and conversation history. When the context window fills up, Codex 'compacts' the conversation — replacing full history with an encrypted summary that preserves the model's understanding.

• Agent loop — why YOU care: Codex doesn't just generate code, it executes a reasoning loop. It reads files, runs shell commands, executes tests, and iterates until the task is done. You get working code, not code that mi...
• AGENTS.md files — why YOU care: Drop a text file in your repo with project-specific instructions (test commands, coding conventions, architecture notes). Codex reads it and follows your rules. Better output without repe...
• Multi-surface architecture — why YOU care: Same agent runs in terminal, VS Code, web, and desktop. Switch contexts without losing state. Your task continues even if you close the browser tab.
• Prompt caching — why YOU care: Conversations grow quadratically (each turn resends all history), but caching keeps computation linear. Long sessions stay fast and affordable.
• Isolated sandboxes — why YOU care: Each task runs in its own cloud container with your repo preloaded. Codex can't access external APIs or your local machine. Security by design.
• Parallel tasks — why YOU care: Assign multiple tasks at once. 'Fix bug in auth, add tests for payments, refactor utils.' Codex works on them simultaneously while you focus elsewhere.
• Bidirectional JSON-RPC — why YOU care: The server can ask for approval mid-task ('run this destructive command?'). You stay in control without babysitting every step.

Who It Is For

If you're a developer who spends time on repetitive, well-scoped tasks like refactoring, writing tests, fixing bugs, or triaging on-call issues — this is for you. OpenAI's own engineers use it to offload work that breaks focus. Not useful yet if you need image inputs for frontend work or want to course-correct mid-task. Best for teams with good test coverage and clear coding conventions.

Worth Exploring?

Yes — if you have a ChatGPT Pro, Plus, or Enterprise subscription, you already have access. The CLI is open source (66k stars) and free to try. The engineering blog posts reveal production-grade patterns for building agents: prompt layering, context management, protocol design. Even if you don't use Codex, the architecture is worth studying. The main gotcha: usage limits on Plus plans have been fluctuating, and cloud tasks take longer than interactive editing.