Graph of Trace lets you watch AI scientists think in real time

Scientific AI agents can now autonomously execute complex research workflows, but their inner reasoning often remains a black box. To bridge this gap, a team of researchers (Gao, Li, Zhao, et al.) developed Graph of Trace, a monitoring and visualization framework accepted to ACL 2026 Demo Track. The system records every fine-grained execution event—including tool calls, API requests, and code executions—and organizes them into a directed graph that updates in real time as the agent works. This structured trace exposes the exact sequence of steps that produced a given result, letting users see not just what the agent did, but how and why it arrived at each conclusion.

To validate the approach, the team conducted evaluations on complex research tasks with domain experts spanning AI, neuroscience, and biology. Experts reported that the structured trace visualization significantly improved their understanding of agent workflows, perceived interpretability, and overall usability for analysis and iterative interaction. The framework addresses a critical bottleneck in human-AI collaboration: the inability to inspect and debug autonomous agents. By making agent workflows explicit as they proceed, Graph of Trace enables researchers to identify where failures emerge, verify correctness, and maintain control over AI-driven scientific discovery.