2026 context-engineering guidance converges on a clear rule: context pollution hurts agent reasoning more than model quality, so naive automatic compaction is the wrong default. Use indexed external memory that retrieves past state on demand, plus rule-based 'context editing' to prune stale tool outputs, so the working window stays small and goal-relevant. SWE-Pruner-style task-aware skimmers that retain only goal-relevant code lines are the emerging coding-agent variant.