What causes context pollution in a long-running agent?

An agent works by adding things to its prompt step after step. Every thought it generates, every tool call it makes, every tool result it reads, gets appended to the running message history. After 30 or 40 steps that history is huge. **Context pollution** is what happens next. The growing pile of intermediate text crowds out the original task or the most recent useful signal. The model still sees everything, but it cannot focus. The original instruction sits at the top buried under tens of thousands of words of detours and the model starts behaving as if the most recent noise is the actual goal. Two things cause it. The first is that nothing is ever removed: tool calls and results stack up forever unless someone prunes them. The second is that not all tool output is useful, so the prompt fills up with content that has no bearing on the next decision. A 50-page document the agent skimmed for one sentence still occupies the full 50 pages in context. The fixes are equally concrete. Summarise old turns into a short paragraph as the conversation gets long. Drop tool outputs that are no longer needed. Keep the original instruction near the front and pinned in a structured place. Some teams run a separate scratchpad agent that holds notes outside the main thread.

**Context pollution** is the degradation of an agent's behaviour as its accumulated message history (tool calls, results, intermediate thoughts) crowds out the original task or the most recent relevant signal. The model still receives all the tokens, but its attention is not uniform, and the more noise it has to wade through, the worse it grasps what to do next. The root cause is that every step appends to the prompt without any compression. A tool that returned a 20,000-token page of HTML stays in the context as a 20,000-token blob even when the agent only needed one fact from it. After 30 steps, the prompt is dominated by transcripts of intermediate work, and the original user instruction sits buried at the top. The failure shows up in two characteristic ways. The first is **lost-in-the-middle**: the model attends more to the start and end of the context than to the middle, so anything important that ended up in the middle stops being followed. The second is goal drift: the most recent noisy tool output starts to look more salient than the original task, and the next action chases the noise. Mitigations are well known. Summarise old turns into compact memory. Drop or truncate tool outputs that are no longer needed. Move long reference material into a retrieval store rather than the live context. Pin the original goal near the bottom of the prompt where attention is strongest.