Introducing Codex

Source: OpenAI Date: 2025-05-16 URL: https://openai.com/index/introducing-codex

Summary

Summary

OpenAI reintroduced Codex in May 2025 as an autonomous software engineering agent — distinct from the original 2021 Codex (a code completion model), this version is a cloud-hosted AI agent that can independently take on software engineering tasks: writing code, running tests, fixing bugs, and submitting pull requests from natural language task descriptions. It runs in a sandboxed environment and works on tasks asynchronously.

Implications

Agentic engineering thread. The 2025 Codex launch is a more significant product than the original because it repositions OpenAI in the software engineering agent race directly against Devin (Cognition) and GitHub Copilot Workspace. The autonomous, asynchronous architecture is the key advance: rather than co-piloting a developer’s session, Codex works on tasks in the background and returns results, enabling a workflow where engineers delegate implementation to AI agents while focusing on higher-level design and review. The cloud sandbox approach also addresses the security concerns of giving an AI agent access to production environments. May 2025 positions this as one of OpenAI’s most direct product-market competition moves against both pure-play AI coding startups and GitHub.