Would automated mutation testing actually make a big difference for agents, or would they just find new ways to game the metric?

I believe it would. Automated mutation testing is the test coverage metric that is nearly impossible to cheat.