In Cisco Outshift's empirical analysis of multi-agent sessions, two cognitive failures caused system output accuracy to decline to 83%