Agentic Path Maps: Exposing an Agent's Traversal Through the Garden of Forking Paths

| Source: arXiv AI

Tags: AI agents, human-computer interaction, overreliance, data analysis, Jeffrey Heer

A user study of 48 people found that showing the alternative paths an AI agent did not take during a data-analysis task reduced overreliance, since users rarely explored those alternatives on their own.

Details

When an agent executes an underspecified task, it makes choices that may conflict with the user's unstated context, but asking for clarification at every step is impractical. The authors propose communicating unexplored paths after execution. They define a design space for presenting alternative paths and study one dimension, the number of alternatives shown, in a data analysis user study (n=48). Participants rarely considered alternative paths unprompted, but overreliance on the AI dropped when alternatives were visualized for them. For teams building agent interfaces, this points to post-hoc path summaries as a way to surface hidden decisions.