The definition of agi keeps shifting - any time an ai can do something, it's just engineering. Current AIs, although narrow, are already superhuman in what they can do. A language AI can converse in more languages than any living human can learn. A chess playing AI can beat any living human. So each time an AI wins on one metric, it's not going to be human level, it'll be superhuman level very quickly.
When an AI finally learns the "only a human can do this" thing, it'll already be superhuman in every other way.