Problems I want to spend the next decade on

How do you give an agent the right context, in the right shape, at the right time, without hand-rolling a system per use case?

What does it look like to instrument taste the way we instrument retention?

The hard-to-detect drifts — directionally correct but wrong answer. What does the eval, observability, and human-review system look like for workflows where mistakes are hard to observe but costly?