Gemini 2.5 Spent Hours Searching for Tea That Didn't Exist in Pokémon Red
The State of Context Engineering
Drew Breunig's AGNTCon talk, "Context Fails & Fixes," traces context engineering from the 2024 "RAG is dead" hype around million-token windows to four persistent failure modes: poisoning, distraction, confusion, and clash. He cites Gemini 2.5 Pro hallucinating a tea quest from Pokémon FireRed while playing the original Red, and a study showing contradictory examples erode instruction-following even in strong models like Claude 4.6 Opus.
Gemini 2.5 Pro at several points was deluded into thinking it had to retrieve the tea in order to progress, and as a result spent many, many hours attempting to find the tea or to give the guard tea.