Agent Trajectory Explorer: Visualizing and Providing Feedback on Agent Trajectories

Michael Desmond, Ja Young Lee, Ibrahim Ibrahim, James M. Johnson, Avirup Sil, Justin MacNair, Ruchir PuriAAAI 2025 · DOI

TL;DR

Agentic systems, which combine LLM reasoning with tool use, generate complex "trajectories" of behavior that are hard for humans to understand. The Agent Trajectory Explorer is a new web-based tool that visualizes these trajectories, making it easier for developers to analyze, annotate, and provide feedback on agent behavior, ultimately improving agent development.

What Problem Does This Paper Address?

Modern AI agents, powered by Large Language Models (LLMs), are becoming increasingly sophisticated. They can now tackle complex tasks by iteratively reasoning, using external tools, and observing the results. This process, however, generates a detailed log of the agent's decision-making, often referred to as a "trajectory." These raw trajectories are essentially long, complex prompts that are difficult for humans to parse, understand, and debug. Without an intuitive way to visualize and interact with this behavior, providing effective feedback and improving agent performance becomes a significant challenge. This paper addresses the need for a tool that can transform these raw, complex agent trajectories into an easily navigable and understandable format for human analysis and oversight.

Methods

The core contribution of this paper is the Agent Trajectory Explorer, a web-based application built with Next.js. The tool's key innovation lies in its ability to visualize agent trajectories. It achieves this through a flexible formatting plugin framework. This framework allows users to convert their agent's raw trajectory data (in JSON format) into a standardized representation that the Explorer can understand and display.

The system supports multiple trajectory formats by mapping a "type field" in the raw JSON to a specific formatting plugin. At runtime, the tool identifies the trajectory type and applies the correct formatter.

The default visualization paradigm is Thought/Action/Observation (TAO), inspired by the ReAct framework. This presents the agent's behavior as a sequence of distinct steps: a thought (LLM reasoning), an action (tool invocation), and an observation (tool output). The tool also offers a raw presentation mode that shows the LLM's inferencing process, providing a lower-level view for deeper analysis.

Furthermore, the Explorer includes a feedback mechanism that allows users to mark specific thoughts or actions with positive or negative feedback.

Datasets

Not discussed in the paper. The paper focuses on the tool's functionality and its ability to process arbitrary agent trajectory data, rather than specific datasets for training or evaluation.

Key Insights and Ideas

  • The "black box" problem of agent behavior: Raw agent trajectories are too complex for direct human comprehension, hindering effective debugging and improvement.
  • Visualization is key to understanding: Transforming sequential, text-based logs into a structured visual representation (like TAO) significantly improves interpretability.
  • Flexibility is crucial for agentic systems: The plugin framework for trajectory formatting acknowledges the diversity of agent architectures and the need for the tool to adapt.
  • Multiple levels of abstraction are beneficial: Offering both a high-level TAO view and a low-level raw output view caters to different user needs and analysis depths.
  • Feedback loops are essential for iterative development: Enabling direct feedback on specific agent steps is a critical step towards more efficient agent refinement.

Key Results

The paper does not present quantitative results from experiments or evaluations. Instead, it focuses on the development and functionality of the Agent Trajectory Explorer. The key "result" is the creation of a tool that:

  • Can ingest and process agent trajectory data from various sources.
  • Visually represents agent behavior using the TAO paradigm (and potentially others in the future).
  • Allows for a basic form of user feedback on agent steps.
  • Provides a more accessible way for developers to understand and debug agent systems.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Limited Feedback Modality: The current feedback mechanism is basic (positive/negative markers). They plan to add richer controls, such as the ability to rewrite thoughts/actions and add commentary.
  • Evolving Visual Paradigms: The TAO paradigm is a starting point, but the authors recognize that other visualization models (e.g., trees, graphs) will be necessary as agentic systems become more complex and adopt new interaction patterns.
  • Complexity of Future Agents: As agent systems evolve (e.g., multi-agent systems, more complex reasoning chains), the need for sophisticated visualization and debugging tools like the Explorer will only increase.
  • Interactive Development: The authors believe that interactive development and debugging are crucial, drawing parallels to chat bot development tools. They envision features like comparing trajectories before and after changes to aid developers.
  • Applying Feedback: Developing effective methods for applying human feedback to agents is a key area for future work.

Summary updated

BibTeX
@inproceedings{desmond2025agent,
  author = {Michael Desmond and Ja Young Lee and Ibrahim Ibrahim and James M. Johnson and Avirup Sil and Justin MacNair and Ruchir Puri},
  title = {{Agent Trajectory Explorer: Visualizing and Providing Feedback on Agent Trajectories}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35350},
}

Loading related content...

Figure