Which of the following are valid strategies for mitigating context pollution in a long-running agent?

**Context pollution** is what happens when an agent's running transcript fills up with stale, noisy, or irrelevant material. As the loop runs, every tool output, every error, and every reasoning trace gets appended. Eventually the useful signal is buried under junk, and the model starts making worse decisions. The valid fixes all do the same thing in spirit. They actively reduce what the model has to read. A sliding window keeps only the most recent steps and forgets the rest. Periodic summarisation rewrites older steps into a compact summary. Selective pruning drops individual entries that fall below a relevance bar. Two of the listed options are traps. Increasing the context window does not remove noise, it just lets more of it accumulate before the agent overflows. The signal to noise ratio keeps getting worse. Disabling tool retries removes one tiny source of clutter, but it also breaks error recovery, which is far more harmful than the small amount of context it saves. The takeaway is simple. Pollution is a content problem, not a capacity problem. You solve it by managing what stays in context, not by making the container bigger.

**Context pollution** is the gradual degradation of an agent's context window as stale tool outputs, errors, and reasoning traces accumulate over a long run. The useful fraction falls, and decisions get worse even though nothing crashed. The three valid mitigations all actively reduce content. A sliding window retains only the most recent N steps. Periodic summarisation compresses older steps every K turns. Selective pruning drops entries that score below a relevance threshold. Each shrinks or filters the transcript so the live context stays mostly signal. The two distractors fail for different reasons. Growing the context window is a capacity change, not a content change. It postpones overflow but lets more noise pile up, so the noise ratio still climbs. Disabling tool retries does remove one pollution source. But it sacrifices error recovery, which matters far more than the handful of tokens it saves. It is not a context management strategy at all. The unifying test is simple. A real mitigation lowers the share of stale tokens the model must read each turn. Anything that merely makes room for more of the same does not qualify.