The interesting thing about Codex now is not that it can write code. That stopped being the main story a while ago. The more important shift is that coding agents are being pulled into the whole development workflow: issues, pull requests, terminals, browsers, docs, remote machines, and review comments.
OpenAI’s Codex for almost everything update is a good example. The release talks about PR review comments, multiple files and terminals, SSH access to remote devboxes, browser-based iteration for frontend work, and pulling context from tools outside the codebase. That is much closer to how software actually gets made than a standalone chat window.
This changes what “using AI for coding” means. It is less about asking for a function and more about directing a worker through a messy environment. The agent needs the repo, the tests, the product intent, the feedback loop, and the boundaries. If those things are missing, a better model helps, but it still has to guess.
That is why I keep coming back to engineering discipline. Agents make good workflows faster, but they also expose vague ones. If the task is unclear, the tests are flaky, and the deployment steps live in someone’s head, Codex will not magically fix the process. It will just make the weak spots more obvious.