We assume AI agents are efficient, but they are actually waiting around for nothing. New research reveals that current large language models struggle significantly when multiple tasks require simultaneous execution with delayed responses. This oversight in evaluating temporal coordination is critical for building truly useful AI systems.

The study introduces AsyncTool, a benchmark designed to test how well AI agents handle asynchronous tool calling. Unlike previous tests that focus on single tasks, this framework simulates real-world scenarios where several heterogeneous tasks run concurrently. It specifically measures how effectively an agent can utilize idle time while waiting for external tool feedback, a key factor in overall efficiency.

The results are stark: delayed tool feedback causes substantial performance degradation in current agents. The research highlights that success depends on a model’s ability to coordinate task switching, track dependencies, and maintain state across multiple workflows. Models that fail to manage these temporal aspects collapse under the pressure of concurrent demands, exposing a major gap in their reasoning capabilities.

The takeaway is clear. To build agents that work in the real world, we must prioritize temporal reasoning and coordination over simple task completion. Future systems need to master the art of waiting and switching, not just executing. Efficiency in multi-task environments is not just about speed; it is about intelligently managing the gaps between actions.