GPT 5.4 has a huge gap in its ability to believe in what's possible, often dumbing down ideas because it's concerned they might not work.
Claude still exceptional in talking through big ideas before getting into concrete details.
Watching Codex try to write a renderer is absolute comedy and frustration. It loves to talk about doing the work, churns for an hour and will often produce results of all the scaffolding around the renderer but not the actual core itself.
The thing I can never get over with Codex is how it likes to talk around something it’s specifically trying to describe. It’s like the smartest idiot sometimes, where Claude just says the concrete idea in plain English.
Who's reinventing git to work seamlessly with multiple agents working in the same repo without the complexity of stashing, branches, work trees, or multiple repos?