Your point about structured context management over raw token dumping matches what I've seen in practice. We run multi-agent workflows where each agent gets a curated 30-50K slice from a shared 200K+ context pool, and the quality difference versus feeding the entire corpus is measurable. The real shift with 1M GA pricing isn't more tokens per request — it's removing the cost penalty for context-heavy orchestration patterns.