AgentWorld Evaluates Agent Reliability with Personality and Adversarial Stress.

Current evaluation methods for agentic information retrieval systems often rely on scripted interactions with uniform users, failing to capture the diversity of human personalities or the brittleness of agents under adversarial conditions. To address this, the "AgentWorld" simulation framework has been developed. This framework integrates user populations driven by the Big Five (OCEAN) personality traits with stateful tool-use environments. AgentWorld employs a comprehensive evaluation approach, including a pass^k consistency metric, structured fault classification, and a dual-control handoff verification system. Crucially, it features an adversarial Risk Analyser that probes agent trajectories for brittleness by branching Monte-Carlo rollouts under various task-aware perturbations. Experiments using conversational analytics and customer support agents demonstrated that personality variations uncover failure modes not visible with uniform testing, such as cross-domain leakage and significant quality gaps across personas. The Risk Analyser also quantified trajectory brittleness, highlighting the dominance of tool and infrastructure-layer attacks.