Context rot is the finding that models do not use a long context evenly. As input length grows, accuracy on the same question tends to fall, and details in the middle of a long prompt are recovered less reliably than details near the beginning or end. The effect has been measured across model families and on simple retrieval tasks, not only on hard reasoning.
The practical consequence is that a large context window is capacity, not quality. Pasting an entire repository or a year of transcripts because it fits can make the answer worse than selecting the three passages that matter. Retrieval, summarization, and explicit pointers ("the clause is in section 4.2") are often more reliable than headroom.
For an agent, context rot compounds: every step adds tool output and history, so a long loop drifts. Trim history, cap tool output, and restate the goal near the end. See context window and estimate the token bill.