How to manage context window in long-running AI agents?

Summarize periodically, use tool outputs (not full responses) in context, trim old turns, use RAG