I can get
eight with EMPATHIC.
be optimal just because I'm not sure you can get more than eight nonadjacent letters. [EDITED to add:] No, you can just about get nine: e.g., QZECTBUMP. But maybe all the nines are obviously impossible?
Let's see. Suppose
there's a 9-letter word. We can cover the 26 keys with the following groups of mutually-adjacent letters: QAW ZSX EDR CFV TGY BHN UJI MK OLP. A 9-letter word will have to have exactly one from each group. (Note: this isn't enough on its own; e.g., ROMANTICS, ALCHEMIST, SHIFTWORK; thank you, Qat.)
Now suppose first of all that
we pick K rather than M; then from OLP we have to take P, and from UJI we have to take U. So we have KPU plus one from each of QAW ZSX EDR CFV TGY BHN. Qat's largest wordlist doesn't find anything matching this other than SPUNKGAVE which I assume is actually a pair of words whose space was somehow missed somewhere in some corpus or other, since it certainly isn't a word. Neither is anything fitting these criteria in a ~400k wordlist I have on my computer.
we have M rather than K; eliminating adjacent letters gives us QAW ZSX EDR CFV TGY BH UI M OLP. Having one of BH rules out G, so we have QAW ZSX EDR CFV TY BH UI M OLP. This still leaves quite a number of possibilities according to Qat, and some of them are even words I know, so let's subdivide according to which we take from TY and on whether we have an A or not. That leaves these cases: [T, no A] QW ZSX ED CV T BH UI M OLP. [T and A] A X ED V T BH UI M OLP. [Y, no A] QW ZSX EDR CFV Y B I M OLP. [Y and A] A X EDR FV Y B I M OLP.
the first (T, no A) has only CHOWTIMES in Qat's union list, and that uses both O and I which are adjacent; the same goes for my list. The other three match no words in either Qat's list or mine.
8 letters does appear to be the best one can do. But this isn't a watertight proof because I haven't checked whether e.g. QRCZYBMIP is in M-W :-).