Isn't Agile more or less gradient descent (well, approximating the gradient for an undifferentiable function), applied to software development?

The objective function is some function of the quality of the product, as measured by the customer, and the work put in. The gradient evaluation is then the process of determining which feature gives the most bang for the buck, implementing it, and repeating.

But gradient descent has problems with local minima. So if that's what Agile is, no surprise it doesn't do surprising projects - even if the customer does know what they want. If the programmers are too inexperienced and it's not just a CRUD, then you'll end up with something like Jeffries' TDD Sudoku solver.