CONTEXT ROT AND CONTEXT BLOAT

SIGNS AND SYMPTOMS OF DEGRADED CONTEXT How can you tell, by watching an agent, that its context has degraded? Doctors diagnose by symptoms, and harness engineers can do the same. Here are common symptoms. First, forgotten instructions. An agent that followed a rule perfectly in a short conversation starts ignoring it after a long conversation. Second, contradictions. The agent says one thing at step ten and the opposite at step sixty, without noticing. Third, repetition. The agent re-does work it already completed, or asks a question the user has already answered. Fourth, missed details. The agent fails to use a fact that appears plainly in the history. Fifth, invented details. When the agent cannot find something, it may fill the gap with a plausible guess, which people call hallucination. Hallucination means the model states something false as if it were true, usually because it is producing likely-sounding text rather than checking facts. Sixth, drifting from the goal. The agent gradually wanders into side tasks and forgets the original objective. Seventh, rising cost and latency with no improvement in quality. When the token count climbs steadily but results do not improve, bloat is a prime suspect. Eighth, errors at the boundary. Sometimes an agent works until the window becomes nearly full and then suddenly fails, because the answer no longer fits. Spotting these symptoms is part of observability, which you will study in Part Seven. Logging the token count at every step makes degradation visible as a graph instead of a vague feeling.