Long contexts are not free: models attend unevenly, recall material in the middle less reliably, and degrade as inputs grow. Agents compensate with compaction, context editing, and careful choices about what to load, and each of those mechanisms decides what the model sees and what it forgets, including safety instructions.
Read in this order
- ArticleEffective context engineering for AI agents (Anthropic)Start here: how to decide what an agent loads, keeps, and drops.
- ArticleContext Rot: How Increasing Input Tokens Impacts LLM Performance (Chroma)Measured evidence that model quality falls as the input grows.
- VideoLLMs for Agent Builders: Context Engineering | Stanford CS329Z Lecture 2 with Michael RyanExplains context limits and structured I/O with lecture chapters. 78 min.