Context Engineering for Engineers

Jeff Huber, the founder of Chroma, discusses context engineering for AI systems, arguing it's a more accurate term than prompt engineering or RAG. He emphasizes that AI is essentially software and the key is determining what information goes into the context window. Huber shares Chroma's research indicating that longer context windows don't always improve performance and that curating relevant information is crucial. He introduces the "gather and glean" model, which involves maximizing recall and then precision to refine context. Huber also explores context engineering in AI agents, highlighting the challenges of managing large prompt histories and the surprising finding that past failure cases can be more beneficial than success cases for agent learning.