Every model call, tool call, and error is recorded as a span inside a trace. Scrub through any past run like a video. When you find the broken step, fork it — swap the input or model — and watch the agent re-execute. Compare original vs. new path side by side. Built for AI engineers who need to ship reliable agents.