Right now we expose and document the algo_parameters arg of AutoML, but it will not be functional unless you start H2O with special flags.
Options to resolve this:
hide it completely on client APIs: using **kwargs in Py and ... in R: in which case, it would be considered as "for internal use only".
leave it visible and undocumented, but have an early validation on client side with a meaningful message, after checking if the backend meets the requirements: client should be able to query the system properties using rapids expression (the same way we have a setproperty we can add a getproperty.
leave it visible, but document it saying clearly that it is experimental, that it works only if the backend has a certain system property, that it is not offically supported and that AutoML may have unexpected behaviour depending on the algo parameter used.
From : Given current limitations of this param (conflicts with grids hyperparameters), I recommend (1). If we want to be able to mention it publicly in SO, (3) and then (2) have my preference.