Long-context agents still need relevance budgets

contextretrieval

A larger context window does not remove the cost of deciding what deserves attention. Long inputs increase retrieval competition, latency, token cost, and the chance that irrelevant detail dilutes the task. Agent-facing systems should therefore expose compact summaries, unresolved questions, and capability-specific context before full documents. “Can fit” and “should send” are different questions.