What I learned analyzing 1k AI agent turns: Sessions are trajectories (unfault.dev)

A recent analysis of 1,000 AI agent interactions led to the development of "Unlost," an open-source command-line interface aimed at enhancing developer interactions with coding agents. It provides three main functionalities: a companion to keep coding agents focused, a recall utility for retrieving context, and a query interface for understanding specific changes. The research revealed that not all conversation turns are equal; while 42% were pivotal (where crucial decisions were made), the remaining 58% were deemed "noise." However, the noise often contained essential context for understanding decision-making, emphasizing the need to analyze all turns for optimal retrieval and intervention. This work signifies a paradigm shift in how sessions with coding agents are understood as trajectories rather than isolated events. By modeling session dynamics, the authors created a stratified intervention system that detects signs of failure before they become critical. The performance improvements stem not only from faster processes but also from a deeper comprehension of interaction patterns. This new approach could reshape how developers, teams, and tool builders perceive coding sessions, highlighting the importance of tracking the sequence of events and contextual signals that predict failure, ultimately improving the efficiency and effectiveness of AI coding interactions.