A highly autonomous system that outperform humans at most economically valuable work.
The biggest problem with this definition is that work ceases to be economically valuable once a machine is able to do it, while human capacity will expand to do new work that wouldn't be possible without the machines. In developed countries machines are doing most of the economically valuable work once done by medieval peasants, without any relation to AGI whatsoever. Many 1950s accounting and secretarial tasks could be done by a cheap computer in the 1990s. So what exactly is the cutoff point here for "economically valuable work"?
The second biggest problem is that "most" is awfully slippery, and seems designed to prematurely declare victory via mathiness. If by some accounting a simple majority of tasks for a given role can be done with no real cognition beyond rote memorization, with the remaining cognitively-demanding tasks being shunted into "manager" or "prompt engineer" roles, then they can unfurl the Mission Accomplished banner and say they automated that role.