Agents Quietly Degrade Before Context Windows Overflow

Long-running AI agents don't crash when context fills—they silently degrade in quality well before hitting the token limit, with reliability dropping sharply at 79% occupancy. This reveals a critical blind spot: monitoring context overflow won't catch the degradation that's already underway. Implementing deterministic checkpoints at ~70% occupancy prevents this silent failure, transforming poor session reliability into predictable, production-ready performance.