큰 컨텍스트 창을 신뢰하지 마라

Vendors' advertised LLM context windows (200k-2M tokens) diverge sharply from effective working capacity, which degrades significantly around 100k tokens—a gap research confirms but marketing ignores. For coding agents consuming tokens rapidly through file reads and debugging, this constraint is mission-critical. The practical solution requires abandoning reliance on auto-compression and instead adopting a structured workflow that preserves information through explicit artifacts and documentation outside live sessions.