BeyondPassAt1-Trajectories

BeyondPassAt1-Trajectories is a large-scale, multi-domain dataset containing 23,760 full execution trajectories from 10 open-weight Large Language Models operating as autonomous agents. Designed to study agent reliability over extended periods, the dataset captures agent behavior across 396 tasks spanning four escalating time horizons (from under 5 minutes to over 2 hours of human-equivalent effort) in Software Engineering, Web Research, and Document Processing domains. Each trajectory logs detailed chronological telemetry—including exact tool calls, environment responses, and token counts—across both standard ReAct and memory-augmented agent scaffolds. Rather than relying solely on binary success metrics, the dataset includes rich diagnostic signals such as Graceful Degradation Scores (partial credit) and Meltdown Onset Points, making it uniquely suited for studying variance amplification, behavioral collapse, and reliability decay in long-horizon AI agent deployments.