Long-horizon tool-calling tasks are a common test for LLM agents, but most benchmarks treat users as cooperative and predictable. A new arXiv paper argues that this misses a critical failure point: how agents handle users who express real-world emotions. The authors introduce τ-Emotion-Bench, a benchmark designed to expose this gap.

In their setup, agents must complete tasks while interacting with users whose emotional states vary and may be negative or volatile. The paper reports that agents can fail even when their tool-use logic is sound, because they struggle to maintain effective interaction under such conditions. This suggests that interaction fragility is a separate and understudied source of failure.

The work highlights a need for benchmarks that go beyond technical correctness and incorporate the messy, affective reality of human users. By doing so, it points toward more robust agent design that accounts for the social and emotional dimensions of interaction.