Look for problems you can test
Many narrow fields have outcomes you can check, or at least check well enough to be useful. Very few people have built that checking into an agent yet.
If an agent can try an answer, test it against a useful signal, and try again, the pattern can work outside math, code, and benchmarks.