Real-world tasks are rarely clean and well-defined; they are often complex and noisy, which poses substantial challenges for LLM agents. Yet existing evaluations of these agents have not provided a systematic framework for jointly characterizing the potential targets and types of perturbations that can occur in such environments.
To address this gap, the paper introduces AgentPerturb, an evaluation framework that applies controlled interaction perturbations to LLM agents. By systematically varying both what is perturbed and how, the framework aims to offer a more comprehensive picture of agent robustness under realistic, noisy conditions.
The work is described in an arXiv preprint (arXiv:2610.21624). While the abstract does not detail the specific perturbation types or experimental results, the core contribution is clear: a structured way to evaluate how LLM agents hold up when their interactions are deliberately disturbed.