Long-running agents can log thousands of steps per run.

Analyzing every step to find what went wrong is slow and expensive.

@kalmiadev has been experimenting with @typesafeai Jev to help choose which steps matter for our harness to analyze.

Same bugs found on real customer