Context Window Budget Management for Long-Running Agents

Strategies for keeping agents within token limits across multi-turn conversations — including sliding windows, semantic compression, and budget guards.