Transitioning from "furthest" to "most reliable" in fitness function

I am trying to train an evolutionary algorithm to take 100 steps perfectly. I figure that in the early game I want to let each AI run once, then select the ones that went the furthest, but in the late game when they start to reach step 100 (no other distinction can be made between them, as step 100 is victory) I want to breed for reaching the end reliably in a series of attempts. What is a good way to implement this? Or is there a better way overall?