One thing you can do is offload from Claude to a dumb local model for summarizing. Local LLM sub-agents.