I agree with this. I am also starting to get a "feeling" of when I can trust an agent and when I can't. Recently I had it throw together a dashboard that displayed some basic linear models based on knobs on the dash, and I didn't really worry about it getting those wrong (I did spot check and it seemed good). But I also had to update a pretty complex flink app with state management changes that it totally borked.

The first task was more self constrained and less production impacting. The latter was detail oriented and required understanding complex distributed systems and state.

I would like to be able to formalize these kinds of tasks. I believe there are lots of confounding variables:

- Access to MCPs

- quality of documentation

- strong existing practices

- examples of similar code nearby

And then we can more easily determine what can be totally handed off and what can't be. I think that last one is most important, but similarly:

- how much this type of algo appears in the training set

Which is maybe part of the "feeling" that we have about what it will do well.