Understanding Context Rot

▤ Summary

This article explains context rot, a production failure mode where LLM output quality degrades as context windows grow longer. It shows how attention gets diluted across more tokens and how important information in the middle of a long prompt becomes harder for models to use. It also highlights practical impacts in coding agents, RAG pipelines, and AI operating costs. The piece recommends compacting context early, using subagents, retrieving fewer but better chunks, and placing critical information at the beginning or end of the prompt.

▦ Linked Companies