Everyone is racing to ship "autonomous agents", but almost nobody is measuring whether these things actually get real work done without babysitting. Here's a practical guide to benchmarking AI agents so you can separate slick demos from systems you can trust in production.