AppWorld-UL: Benchmarking Diverse Agent-User Interactions for Tool-Use

Tool-use agents that address day-to-day digital tasks such as ordering groceries must not only operate applications, but also interact with the user, e.g., to ask clarification questions, prompt for confirmation, and inform the user when the instruction is infeasible. However, current benchmarks for evaluating agent-user interactions do not capture the diversity of such interactions. Further, they operate in small environments with few, often non-state-changing, APIs. To address this gap, we introduce AppWorld-UL, a ``user-in-the-loop'' benchmark of 516 challenging tasks requiring diverse agent-user interactions. Building upon the AppWorld framework with 9 popular simulated apps like Amazon and Spotify, we systematically modify original tasks to introduce ambiguities and constraints that necessitate various types of agent-user interaction. User behavior is simulated by an LLM prompted to respond with carefully designed knowledge boundaries, offering more reliable simulation than the unconstrained or overly rigid alternatives used in prior work. Our evaluation reveals that a state-of-the-art LLM, Claude Opus 4.7, achieves only 48.6\% success on AppWorld-UL, and only 35.7\% on the harder, compositional subset. On the stricter, scenario-level metric, compositional task performance drops to only 21.3\%. Our analysis reveals that correct user-interaction is crucial for success. This demonstrates the benchmark's difficulty and its potential to advance research on user-in-the-loop tool-use agents.

AI agents are increasingly expected to help people with everyday digital tasks, such as ordering groceries, managing playlists, or booking services. To do these tasks well, they need more than the ability to click buttons or call software tools: they also need to communicate with users when instructions are unclear, risky, or impossible to complete. In this work, we introduce AppWorld-UL, a user-in-the-loop benchmark designed to test how well AI agents can work with users while using digital applications. AppWorld-UL contains 516 challenging tasks built on simulated apps such as Amazon and Spotify. We create these tasks by modifying existing digital tasks so that agents must ask clarification questions, request confirmation, or explain when a request cannot be completed. We also design a simulated user that can answer agent questions within carefully controlled knowledge boundaries, making the evaluation more reliable. Our experiments show that even strong AI models still struggle with these user-in-the-loop tasks, especially when multiple interaction challenges appear in the same task. These results suggest that future AI agents need better abilities to recognize when they need user input and to ask for it effectively.