AMG iterative coarse grid solver has a hard-coded high tolerance

If no direct solver backend is found, AMG uses BiCGSTABSolver, but with a hard-coded tolerance of 1e-2. This leads to different solutions in one of the Dumux tests, depending on whether a direct solver is available or not, although all other solver parameters are the same. Is there a reason why such a high tolerance is expected to produce a good enough approximation?

In any case it would probably be nice to be able to pass the coarse grid solver tolerance and iteration number from outside similar to the smootherArgs. Any hints on what has to be done to implement that cleanly?