Abstract
Tool-use agents that handle day-to-day digital tasks, such as ordering groceries, must not only operate applications but also interact with users—for instance, to ask clarification questions, prompt for confirmation, and inform users when instructions are infeasible. However, existing benchmarks for evaluating agent-user interactions fail to capture the diversity of such interactions and typically 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 consisting 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 a large language model (LLM) prompted to respond within 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.
Blogger's Review: AppWorld-UL provides a significant standard for evaluating agent-user interactions, revealing the limitations of current LLMs in complex human-computer interactions. This not only advances the research on tool-use agents but also offers valuable directions for future improvements. By introducing ambiguities and constraints, researchers can better understand and optimize the performance of agents.