Atlas helps people understand long agent runs and decisions
Atlas: Making Sense of Long Horizon Agent Trajectories
Human-Computer Interaction
Summary
AI agents can now run long, complex tasks like researching or experimenting, but people find it hard to understand what the agent did afterward. The authors created Atlas, a tool that shows the agent’s entire sequence of choices as a branching tree and displays the important results in different ways. In a test with 14 people, Atlas helped users see what decisions the agent made, what results it produced, and how everything connected to the goal, all with less effort than a standard tool.
What this means in practice
- •For ai developers: Visualize and interpret the sequence of decisions and results from complex AI agent runs to improve debugging and refinement.
- •For data analysts: Review outputs and pathways of automated data exploration tools to verify findings and connect outputs to goals more clearly.
Authors
Sangho Suh, Joseph Chee Chang, Pao Siangliulue, Amy X. Zhang, Matt Latzke, Jonathan Bragg, Doug Downey, Daniel S. Weld
Abstract
AI agents can now search literature, generate hypotheses, analyze datasets, and run experiments at a scale difficult for researchers to match. Yet scaling scientific execution does not automatically scale researchers' ability to make sense of what agents produce. When researchers return to a long-horizon run, they must evaluate work they did not monitor: what the agent produced, which decisions shaped those outputs, and whether those decisions served the research goal. We frame this post-run sensemaking as a representation problem across three linked levels: goal, decision, and artifact. From this framework we developed Atlas, an agentic workspace with a trajectory tree, which renders the run as a branching structure showing what the agent chose and set aside, and artifact projections over six representations. In a within-subjects study (N=14), participants reported higher decision transparency, artifact sensemaking, and goal connection with Atlas than with a standard baseline, together with substantially lower subjective workload.