You still have to double check the Agent's grading as they still make frequent minor mistakes, that while they don't really impact coding as much or can be caught with tests, that is not the case here.
(That is basically the argument for OCR in the postal system -- it does make mistakes but it makes fewer mistakes than human beings.)
Really though, this is missing the point. Yes, grading tests and homework is different from sorting mail, but both are tedious tasks that involve a lot of repetitive actions -- hard for the human brain, but easy for software. Software does not get tired or distracted by other responsibilities (or non-responsibilities). We have already been using software to grade tests for a long time now, but mostly for multiple choice questions; AI will expand the scope of that.