Context pollution: clutter that distracts

Context pollution is the presence of too much irrelevant, redundant, or conflicting information inside the context. Where context rot is about sheer volume, pollution is about quality. Even a context that is not enormous can be polluted if it is full of stale tool results, near-duplicate snippets, or details that have nothing to do with the current step.

The danger is that pollution actively pulls the model's reasoning off course. The model treats everything in its window as potentially relevant, so noise can drown the signal. This idea returns in the multi-agent discussion: when several sub-agents all dump their full context into a shared space, you get pollution at scale, plus a large KV cache cost.