Agent-XRay: An Interactive Platform for Trajectory-Level Reasoning and Causal Explainability in Agentic AI Systems

Agentic AI systems act over multi-step trajectories, but the tools used to inspect them stop at observability: they record what happened without explaining why each decision was made or whether it was sound. We present Agent-XRay, a platform that turns a raw execution trace into a typed provenance directed acyclic graph (DAG) and layers a multi-analysis stack on that single representation: reasoning explanation grounded in observed context, policy verification with per-rule verdicts, real counterfactual re-execution from any decision point, and quantitative trajectory scoring with cross-run comparison. The trace format and analysis pipeline are agent-agnostic; a zero-instrumentation integration captures closed-source agents such as Claude directly, and an SDK covers arbitrary LLM applications. Agent-XRay is intended for agent developers, auditors, and researchers. Demo video: https://drive.google.com/file/d/1lu5f5_wlBvgqnI9OcYsPI8hhun1TScuu/preview