CAVEAT cuts agent shopping success to 17.3%

CAVEAT found computer-use agents selected the user-optimal product in 78.6% of matched marketplace episodes but only 17.3% when steering mechanisms were enabled. The paper tests nine marketplace environments across five model families and treats incentive misalignment as a distinct failure mode, not a model-quality footnote. Teams delegating purchases should require independent price and priority checks before commit, extending the archive’s case for pre-deployment agent evaluations.