Put agent harnesses
on the same track.

Launch one controlled turn across six runtimes. Compare the answer, event trail, latency, usage, and failure shape without changing the task.