Reading Right, Acting Wrong: Diagnosing Decision Failures of System One Models in Agent Loops