Comment by momobassit

That is actually quite similar to my experience using multiple AI coding agents, including Codex and Claude Code. There is an initial phase where things go very well, but then things start to get slower and I feel like I'm stuck in a loop trying to get the agents to fix things without breaking anything else.

I find them most useful for making prototypes to show clients who are unimpressed with a presentation/document, but I end up doing most of the implementation myself. Which is fine.