Alexander Embiricos: How Codex grew 20x by moving into IDEs

Codex’s vision is a software engineering teammate that participates end‑to‑end—from ideation and planning, to implementation, validation, deployment, and maintenance—acting on its own where safe, instead of waiting for explicit prompts.

OpenAI has learned that models are far more effective when they can use a computer, and the best way to do that is by writing code (e.g., via shells and sandboxes), making code execution a core competency for any serious agent, including non‑developer use cases.

Codex’s early cloud‑only, asynchronous model was “too far in the future” for most users; shifting to an IDE extension and terminal‑based agent that runs in a sandbox on the developer’s own machine massively simplified onboarding and unlocked explosive growth.

As agents write more code, engineers increasingly spend their time reviewing AI‑generated changes and ensuring safety and correctness, which is often less fun; Codex is now focused on better automated testing, validation, and code review support to relieve this pressure.

Features like long‑running workflows and “compaction” (moving work across context windows) only work well when the model, the API, and the tool harness are co‑designed—OpenAI’s tight product–research integration is a key advantage in building robust agents.