Detecting Perspective Shifts in Multi-Agent Systems

Generative models augmented with external tools and update mechanisms (or agents) have demonstrated capabilities beyond intelligent prompting of base models. As agent use proliferates, dynamic multi-agent systems have naturally emerged. Recent work has investigated the theoretical and empirical properties of low-dimensional representations of agents based on query responses at a single time point. This paper introduces the Temporal Data Kernel Perspective Space (TDKPS), which jointly embeds agents across time, and proposes several novel hypothesis tests for detecting behavioral change at the agent- and group-level in black-box multi-agent systems. We characterize the empirical properties of our proposed tests, including their sensitivity to key hyperparameters, in simulations motivated by a multi-agent system of evolving digital personas. Finally, we demonstrate via natural experiment that our proposed tests detect changes that correlate sensitively, specifically, and significantly with a real exogenous event. TDKPS is the first principled framework for monitoring behavioral dynamics in black-box multi-agent systems -- a critical capability as generative agent deployment continues to scale.

Modern AI systems increasingly act as "agents" that combine language models with tools, retrieval, and external data, and their behavior can drift over time. Detecting this drift is challenging when only an agent's inputs and outputs are observable. We introduce the Temporal Data Kernel Perspective Space (TDKPS), a framework that embeds agents into a low-dimensional space capturing their relative behavior across time, paired with statistical tests for detecting change at the agent and group level. To assess whether TDKPS detects real behavioral change, we constructed a natural experiment: 99 AI agents grounded in U.S. Congresspersons' tweet histories, queried over seven years on public health, and two progressive control groups: general politics (which would correlate with latent shifts in online persona behaviors), and candy and chocolate (which serves as a global control). Detected shifts clustered around the onset of COVID-19 and were concentrated specifically in public health queries, with no similar pattern in the control topics. Because COVID-19 emerged independently of congressional behavior, this topic-specific, temporally aligned signal supports the conclusion that TDKPS detected genuine behavioral changes mirroring documented shifts in political opinion during the pandemic. As agentic systems proliferate, principled monitoring will be essential for trust and safety. TDKPS is a first step toward doing so without requiring internal access.