Researchers have developed ChainWorld, a system that composes atomic tasks into long-horizon desktop workloads. This is done through directional compatibility search, which preserves the original task evaluators. The resulting workload consists of chains of tasks with lengths two to four, and is evaluated in both single-turn and multi-turn settings. Results show that current computer use agents struggle to complete these tasks, with maximum chain completion rates at 31%. Multi-turn evaluation improves performance for three models, but still poses challenges. --- Why it matters: This matters because it highlights the limitations of current AI systems in handling complex desktop tasks that require sustained state across multiple objectives. Understanding and addressing these gaps is crucial for developing more realistic and effective AI agents. Source: https://arxiv.org/abs/2606.21654