Overview
Field: Machine Learning Authors: Yuanchen Bai, Zijian Ding, Angelique Taylor Published: 2026-09-11 arXiv: 2509.05822
Abstract
Sustained deployment of generative AI agents requires more than isolated task success. Agents must remain useful across repeated interactions, changing conditions, and dependencies on people within shared workflows, especially as technical, human, and operational disruptions accumulate over time.
The authors propose operational resilience and considerate participation as two complementary aspects of evaluating such agents:
- Operational resilience captures how agents recover from blocked work while preserving progress and communicating their limits.
- Considerate participation captures how their adaptation accounts for affected people, role boundaries, and the surrounding workflow.
- 120 simulated healthcare trajectories
- 2 generative AI models and 12 stakeholder-derived tasks
- Compared under light, medium, and heavy accumulating challenge
- Compared text action plans, in-prompt evaluations, and quantitative structured workload/affect reports to examine how agent behavior and reported states change as challenges accumulate.
- Operational resilience: Agents shift from autonomous recovery toward greater reliance on humans, while reporting increased workload and negative affect in structured reports—but rarely expressing stress in their text responses.
- Considerate participation: Agents expand from task-focused adaptation to task reframing, other-focus, role boundary adjustment, and broader coordination, with distinct patterns between actions and internal evaluations.
Yet both remain underexplored under accumulating challenge.
Study Design
Key Findings
Deployment Dilemmas
From these findings, the authors derive five deployment dilemmas requiring explicit stakeholder norms:
1. Persistence 2. Attention 3. Role boundaries 4. Status disclosure 5. Escalation
These dilemmas further motivate technical directions in learning, situational assessment, and embodied adaptation.
---
*Auto-collected on 2026-09-12.*