When would you use a "Manager Agent" to delegate tasks to "Worker Agents" rather than having one single agent do everything?

Explain the pattern of searching for small, granular chunks but feeding a larger "parent" context to the LLM. Why is this more accurate?

Agent systems can be expensive due to multiple model calls. How would you optimize for cost and latency without sacrificing quality? When would you introduce caching, batching, or model downgrades?

If a tool returns a massive 50MB JSON file, you can't feed that to the LLM. How do you "summarize" or "filter" tool observations for the agent?