I’ve been using Codex for a while, and a few of its problems get in the way enough that I want to write them down.
It agrees with the user too easily
When you give it a request, it assumes what you said is right. It doesn’t stop to think about whether something is missing or whether the request has problems. It just starts working.
But a lot of the time, the request itself is exactly what should be talked through first. A good collaborator asks a few questions or points out what’s unclear, instead of taking everything at face value.
It’s bad at correcting itself
While working on a task, it rarely looks back to check itself. It may have misunderstood the request from the start, yet it can keep going for a long time without once stopping to ask you or check in.
Only when it’s done do you find out the result isn’t what you wanted at all. Most of that time was wasted.
The UI doesn’t tell you what it’s doing
The Codex agent isn’t actually slow. The problem is that it hides a lot of what it does. It might be running a Bash command or reading some context, but none of that shows up. You just see “thinking” sitting there, unchanged.
When you can’t tell what it’s doing in the background, it feels slow and makes you impatient. Speed isn’t the issue. It just doesn’t tell you anything.
Where it’s strong
To be fair, Codex does have a lot of features, and its Computer Use is much stronger than Claude’s.