ChainWorld: Composing Long-Horizon Desktop Workloads from Atomic OSWorld Tasks

ChainWorld introduces a framework for composing long-horizon desktop workloads from atomic OSWorld tasks, addressing the gap in evaluating computer use agents beyond single atomic tasks. It features 347 task chains of varying lengths and compares single turn versus multi turn evaluation methods, revealing a maximum chain completion rate of 31%. This work is significant for practitioners as it highlights the complexities of task management and state preservation in AI agents, informing the design of more robust systems for long-duration user interactions.