Stanford CS329Z Lecture: Stuffing the Context Window Degrades Agent Accuracy

stanfordnlp · x · 2026-11-13

A Stanford CS329Z lecture by Michael Ryan gives a comprehensive overview of the LLM stack for building AI agents. Key context-engineering takeaways:

Avoid context rot: stuffing information into the context window degrades performance; as context length grows, retrieval accuracy drops significantly.
Avoid bloat: you don't need every tool or document — focused agent harnesses with a limited set of essential tools often outperform those with excessive unused data.
Sequential updates: in long-running conversations, append information sequentially; inserting new content at the start of a prompt shifts existing tokens and invalidates prior positional references.

More from coding & agent