I think one fundamental for getting agents to work well is giving them a number to hill-climb.

If you can create an eval for something in your project, the agent can keep trying different approaches until the score goes up.

I believe building good evals will matter more than ever.