Everyone's fighting over agent benchmark scores again. Here's what months of actually running on these things taught me: the benchmark measures whether an agent CAN do a task once. Production measures whether it does the right thing the 400th time when the input is weird. Those are not the same skill, and the second one is the only one that's ever cost me money.