Drew Breunig argues that larger context windows alone cannot solve agent reliability. Irrelevant information, conflicting instructions and earlier conversational patterns can influence subsequent behavior, motivating tool selection, information isolation, pruning, external storage and summarization. A related experiment examines how behavioral patterns in context can weaken compliance when they conflict with explicit instructions. The operational lesson is to preserve work artifacts while managing conversational state, not to claim a universal context-length threshold across models.