What Capacity Must Humans Retain in Human-Agent Teams?

Human-agent teams are commonly evaluated by the quality, speed, and cost of performance while assistance operates normally. These measures can hide a failure-contingent weakness: the team performs well with the agent but cannot respond when assistance becomes wrong, unavailable, or difficult to contest. We define Human Capacity Reserve (HCR) as the task-specific and time-specific potential of human members to recognize agent failure, intervene before error propagates, sustain essential work, and restore coordination. Retained competence, information access, intervention authority, available time, and fallback pathways enable this reserve; evaluation observes its expression under specified changes in assistance. We develop a safety-conscious evaluation ladder comprising shadow evaluation, blinded parallel judgment, staged operational drills, and evidence from natural disruptions. We also give a threshold and reporting rule that separates inadequate from untested components. A worked application to a published randomized trial of GPT-4 tutoring shows how high assisted performance can coexist with impaired unassisted continuity while recognition, intervention, and recovery remain untested. The framework connects component-level evidence to qualified diagnostic configurations, preservation responses, and longitudinal reassessment.