Yeah it’s clever, but doesn’t feel overly agentic. It stops prematurely and constantly requires persuasion to continue and do the thing. I know the new harness has added explicit instructions to make it behave as we might typically expect, but it doesn’t seem to be enough!
“If you want me to push ahead…” this is after 3 interventions requesting it do the thing, and giving it permission to do so - it’s still checking with me that it’s ok to proceed.
The output is very dry and verbose, and not human/user-friendly. e.g. I asked for a comparison of 2 features, and it produced walls of text, rather than a comparison table.
I am using it because it’s free, but I don’t feel like I want to use it.
codex needs some persuation sometimes, like using a tool it has. Very seldom IME.
I use Codex CLI all the time. In Cursor it’s much faster, they are probably hosting on Cerebras.
Initial impression was way perfect. Today it feels nerfed already - less intelligent, less instruction following. Codex-high in CLI is better than todays experience, but very slow.
Today I did “real work”, not sure that is the difference? It still implements well. To compare: I wouldn’t let Opus 4.5, or Gemini 3 Pro high touch code (while it is REALLY good for finding bugs).
So it’s more prompting then using codex cli - but much more fun since fast and free. And it does not poison your code like Opus or Gemini do.
