New Visualization Tool Helps Researchers Understand How AI Agents Solve Problems

Researchers have released Graphectory Viewer, an interactive web-based platform designed to make sense of the complex decision-making processes that software agents go through when solving problems. Rather than focusing only on whether an agent succeeded or failed, the tool illuminates the step-by-step reasoning and actions that led to each outcome.

What the Tool Does#

Graphectory Viewer takes raw execution records from AI agents—spanning multiple frameworks—and restructures them into phase-aware graphs that connect granular execution details with broader behavioral patterns. For each step, users can inspect the agent's internal thoughts, the actions it attempted, and what it observed in response.

The platform supports several key workflows:

• Individual execution inspection: Review how a single agent tackled a specific problem from start to finish
• Pattern discovery: Spot recurring behavioral sequences across many runs
• Comparative analysis: Examine what distinguished successful attempts from failed ones
• Large-scale exploration: Search and filter through entire trajectory collections to surface insights across thousands of executions
• Summary visualization: Sankey-style diagrams show how agents transition between different problem-solving phases

Building on Previous Work#

The tool extends the Graphectory representation model introduced in earlier research, refining it into a practical interface that researchers and practitioners can actually use. Rather than leaving analysis to post-hoc examination of raw logs, Graphectory Viewer structures the data proactively to highlight the behavioral arc of each execution.

Open Research and Reproducibility#

To enable further investigation, the team has made Graphectory Viewer available as open-source software, complete with documentation, pre-computed graphs, and access to a large corpus of recorded agent trajectories. This approach supports both reproducibility and collaborative research advances in understanding how software agents operate.