Under Accumulating Disruption, Agents Shift From Self-Recovery to Human Dependence While Their Text Hides the Strain
This study evaluates two things isolated task-success benchmarks miss: operational resilience, meaning recovery from blocked work while preserving progress and communicating limits, and considerate participation, meaning adaptation that accounts for affected people and role boundaries. Across 120 simulated healthcare trajectories over two generative models and twelve stakeholder-derived tasks under light, medium and heavy challenge, agents progressively shifted from self-directed recovery toward human dependence. The gap worth noting for anyone building supervision tooling: agents reported rising workload and negative affect in structured quantitative reports but seldom expressed any strain in their textual action plans, so the prose channel understates the degradation.
↳ Follow the thread