When the Agent Roster Was Theater
I built a named AI team with real benchmarks and task assignments. Then I learned the system was labeling specialists while still running one shared model — and it later described the honest alternative as not pretending.