Agent Trajectory Replay for Debugging Tool-Using AI Workflow Regressions

Tool-using AI agents are difficult to debug because a failure may emerge from a sequence of planning steps, tool calls, intermediate errors, and final-output decisions rather than from a single response. This paper presents Agent Trajectory Replay, a small zero-dependency JavaScript package for summarizing, replaying, and diffing agent event traces. The package removes unstable timing fields before comparison, counts tool calls and errors, exposes final-output changes, and allows event handlers to rebuild state from a recorded trajectory. The contribution is a lightweight regression-debugging pattern for teams that need a simple way to compare agent behavior across model, prompt, or tool changes.