Three rules are available:

1. Always use H. Simple, but you overpay on tasks that L could have handled.

2. Always use L. Cheap, but you fail on tasks that need H.

3. Run both in parallel, take whichever works. Highest completion rate, but you pay for redundant work even when one agent alone would have sufficed.

-> 3 as described is so weird conceptually (presumably if H is the strongest model, this strategy would be more expensive than 1 in all cases, with not quality improvement) -- do you mean use L and then H if L failed?