Context Window Overflow Is a Silent, Measurable Failure Mode in AI Agents

AI agents in production can silently truncate long prompts — such as multi-message threads with logs and stack traces — and still return confident, fluent answers based on incomplete information. Most agent frameworks drop overflow content without raising errors, leaving evaluations green while users receive responses built on a partial view of the problem. The core issue is that teams typically assess output quality rather than verifying whether all required evidence actually reached the model. The author argues this is a Tier 1, deterministic failure — provable by inspecting the resolved prompt's token count and content — not a subjective quality issue suited for model-as-judge evaluation. The recommended fix is to capture and measure the fully assembled prompt before each model call, gating on hard token limits rather than relying on downstream output review.