My team worked with OpenAI in late 2021 / early 2022. That time we only had GPT-3 available and we were working on a chatbot for healthcare. We ended up going down in a rabbit hole and building all sorts of things around GPT-3 to make it work reliably as a chatbot. We didn't succeed and the OpenAI team recommended us to wait for their new model (that became the core of ChatGPT a few months later). I see similarities today with agents built with GPT-4. As GPT-3 wasn't built for chat, GPT-4 wasn't built for agents. As trying to make GPT-3 to perform well for chat applications was a mistake, trying to make GPT-4 to perform well for planning and agentic behaviour is a mistake. I expect GPT-5, whenever it comes out, to be a step change for agentic applications and for this reason working on "agents with GPT-4" doesn't seem like a viable long-term strategy to me.