Retrieval quality defines AI agent architecture

Agentic systems fail when context-building fails, not when the LLM falters. A client named Specstory wanted users to query agent history for decisions like why Authlib was chosen for authentication. Improving generation models cannot fix systems that cannot find the right sources first.