TraceGraph: Shared Decision Landscapes for Diagnosing and Improving Agent Trajectories

In three linesTraceGraph is a graph-based framework that transforms multi-model agent trajectories into shared decision landscapes. It builds graphs over state-action-observation spaces, identifies productive cores and trap regions, then proposes a trap-aware recovery pipeline. On SWE-bench, this approach improves resolution rate from 40.4% to 43.5%.

Summary generated by Claude — human-verified