The “model spends; harness budgets” framing resonates, especially because context is both a quality variable and a billable resource. I’d make the budgeter layered: first remove deterministic repetition or low-value structure locally, then use retrieval or model-based summarization only when the task actually needs semantic selection. I’ve been experimenting with Contextpress (GitHub - Taha-azizi/contextpress: Deterministic context compression for LLM chat, RAG, and agent pipelines · GitHub · contextpress · PyPI) as a deliberately boring deterministic first pass over agent/tool message lists. It seems easier to reason about in production because identical input produces identical output, and the evaluation can focus on success rate, input tokens, and recovery behavior rather than trusting a claimed compression percentage.