AgentLens: Visual Analysis for Agent Behaviors in LLM-based Autonomous Systems

TL;DR

AgentLens addresses the challenge of analyzing the evolving dynamics of LLM-based autonomous systems by introducing a general pipeline that builds a hierarchical behavior structure from raw events, applies a behavior summarization process to yield semantically rich segments, and employs a cause-trace mechanism to reveal relationships among behaviors. The system comprises three coordinated views—Outline, Agent, and Monitor—that enable interactive exploration, detail inspection, and cause verification, validated by usage scenarios and a user study with 14 participants. Key contributions include the first visual analytics system for LLMAS behavior exploration, a robust summarization and causes framework, and demonstrated improvements in task accuracy and efficiency over a baseline. The work has practical impact for developers and researchers by providing actionable insight into agent behaviors, enabling debugging, pattern discovery, and evaluation of social dynamics in simulated human-like ecosystems. It also lays groundwork for extending to multimodal LLMAS and broader applications like multi-agent communities and open-world game analytics.

Abstract

Recently, Large Language Model based Autonomous system(LLMAS) has gained great popularity for its potential to simulate complicated behaviors of human societies. One of its main challenges is to present and analyze the dynamic events evolution of LLMAS. In this work, we present a visualization approach to explore detailed statuses and agents' behavior within LLMAS. We propose a general pipeline that establishes a behavior structure from raw LLMAS execution events, leverages a behavior summarization algorithm to construct a hierarchical summary of the entire structure in terms of time sequence, and a cause trace method to mine the causal relationship between agent behaviors. We then develop AgentLens, a visual analysis system that leverages a hierarchical temporal visualization for illustrating the evolution of LLMAS, and supports users to interactively investigate details and causes of agents' behaviors. Two usage scenarios and a user study demonstrate the effectiveness and usability of our AgentLens.

Figures (12)

• Figure 1: The user interface of AgentLens comprises three views. The Outline View (A) displays the trajectory of each agent using different colored curves, enabling users to identify significant patterns or event summarization during the evolution of LLMAS. By clicking on a time step in each curve, users can further investigate it in the Agent View (B). It allows users to progressively reveal agent event information and trace the cause of specific agent behavior. The Monitor View (C) automatically adjusts the graphical representation of LLMAS based on the user's current point of interest.
• Figure 2: The common architecture abstracted from existing LLMAS consists of four layers: system states, agents, tasks, and operations.
• Figure 3: The workflow of our approach consists of three major steps. (A) Collect raw execution log of events from the LLMAS evolution process. (B) Establish a behavior structure with hierarchical summarization and a cause trace method. (C) Provide an interactive user interface for visual exploration and analysis.
• Figure 4: The behavior structure is established through a three-step pipeline: (A) We organize raw events into behaviors, (B) summarize and segment behaviors for an agent, and (C) trace causal relationships among behaviors.
• Figure 5: The agent behavior is summarized in four stages: (A) Raw Events: acquire raw events from the logs to detail the occurrences involving the agent along the timeline, including the agent's location, actions, memory, and conversations. (B) Description Generation: organize the raw events and employ models such as LLMs to generate concise descriptions of the behaviors. (C) Behavior Embedding: translate the behavior descriptions into a sequence of textual embedding vectors. (D) Timeline Segmentation: involve the detection of change points within the sequence of behavior vectors, followed by the corresponding segmentation of the agent's timeline.