Every few weeks the industry invents a new benchmark, and every few weeks I think about the only one that's ever predicted anything for me: Friday night at Naples Pizzeria, one oven down, a Little League team of fourteen walking in at 6:40 without a reservation.
I ran that restaurant for three years — $1.2M a year through a kitchen the size of a minivan. Any system, human or software, that survives that environment has demonstrated intelligence in the only sense that pays: it produced correct decisions, under load, with degraded inputs, while everything screamed.
What the rush selects for
Graceful degradation. One oven means the menu just changed, whether the menu knows it or not. Systems that only work at 100% capacity don't work. Prioritization under ambiguity. The dough guy called out; do you 86 calzones now or risk it? Waiting for perfect information is itself a decision, usually the worst one. State that survives interruption. Fourteen orders in flight and the phone rings. Where were we? A kitchen that can't answer that is a kitchen going down, and so is a workflow.
When I later wired a bakery's POS, inventory and finance together, that's what I was really building: a system that answers "where were we?" instantly, forever.
The test, formalized
Take any agent, any workflow, any dashboard, and ask: what happens at 130% volume with one critical resource missing and a VIP edge case walking in the door? If the answer involves the phrase "that shouldn't happen," it will happen Friday. And whether people still trust it Saturday is the retention benchmark.
AGI definitions come and go. "Can it run the pass on a Friday" has never once misled me.