Context Degradation

If you've ever watched an agent slowly lose the plot over a long conversation, this explains why. It covers five degradation patterns: lost-in-middle (that U-shaped attention curve where middle content gets ignored), context poisoning (errors that compound through repeated reference), distraction (irrelevant info eating your attention budget), confusion (model can't tell which context applies), and clash (contradictory information derailing reasoning). The practical bit is knowing these are predictable, not random failures. When your agent starts hallucinating after 50 messages or calling the wrong tools, you're probably hitting one of these patterns. Useful for debugging production issues and understanding why careful context management actually matters at scale.