Context Poisoning
Context poisoning is when bad, stale, or excessive information in a model's context window degrades its reasoning — the agent gets buried in irrelevant tokens or misled by wrong data, and its answers get worse even though the model is fine.
Also known as: context pollution
More context isn’t better context. As an agent runs, its window fills with tool output, prior turns, and retrieved chunks — and a lot of it goes stale or was never relevant. Context poisoning is what happens next: the signal the model needs gets buried in noise, or it anchors on wrong data, and the quality of its reasoning drops. The model didn’t get worse; its working memory got polluted.
The fix is active context management: pull only what’s relevant, drop what’s gone stale, summarize or compact long history, and watch the token budget. It’s a core reliability problem for long-running agents, and a big part of why context engineering exists as a discipline.
