But how is it even possible (bad harness?), or wise, to send 500k or 1M tokens per call? Regarding cache, how are you not hitting the 1hr cache? Also, start new chats early and often!

I have been "agentic coding" since Sonnet 3.5 and after this paper came out, it became my bible:

https://github.com/adobe-research/NoLiMa

Last I checked, all models suck as you fill the context window. "Context engineering" is how you do this whole thing.