Why would would Agents need multiple turns and a framework like Pythagora in a (near-future) world with GPT4-level of capability and 10M+ token context?
Imagine GPT4 (or any other LLM) as a very eager but not very bright junior developer that just started to work with you. It's good, but it'll need a lot of situational management for it to not go wildly off the track.
What improved models will bring us in the future is making Pythagora and other such tools work better for large and more complex projects. The tipping point will come when for example you'll be able to load Pythagora in Pythagora and continue development. While we do build some auxiliary/external tools with Pythagora, the core is still handcrafted mostly by a human, and I'm pretty sure that's the case with other code gen tools as well.
Because it's better. This was the insight behind AlphaGo, AlphaCode and AlphaGeometry. In Go, trying to get the entire solution in one forward pass required a 1000x larger neural net to achieve the same performance according to Noam Brown.
Intelligence requires iteration, introspection, planning, sub-task breakdown, and trial and error with feedback from reality (error messages, print statements). These meta reasoning structures (whether hard-coded or, ideally, learned) will move LLMs closer to how humans reason.