There have been lots of draws so far in the Season 8 Superfinal of the Top Engines Chess Competition (TCEC), but Komodo has managed to win three games while Stockfish has won but one.
The last game, game 38, was won by Komodo, and it was an impressive victory. That said, it also revealed one of the things engines still struggle with this, or - perhaps more precisely put - still cannot do: recognize in some conceptual way that a position is drawn. Failing to recognize a fortress as a fortress is the most familiar manifestation of this shortcoming, but not the only one. Pull up the game from the file structure (Season 8-Superfinal-game 38) on the archive page, and look at the position before White's 87th move. The plan with 87.Rf3 followed by 89.h4 is horrible, and practically speaking, it would be horrible even if it somehow turned out to be a draw.
Before striking on this "opportunity", Stockfish (and Komodo too) seemed to focus on lines where White kept his rook active on the queenside, being sure to be able to defend the d-pawn. In those scenarios Black would almost surely have to play ...R~3 followed by ...Nxh3(+) Bxh3 Rxh3 at some point, and the resulting endgame would be drawn. (If I'm not mistaken, it wouldn't even be a difficult draw.) A human seeing this would just stick to the program, steadfastly ignoring rabbit trails that might in some utterly meaningless sense be - or in this case, merely seem, briefly - one tenth of a pawn better. And the human would be right: there isn't anything to calculate or figure out; the work has been done and the half point is there for the taking.
This is not some sort of quirk that only pertains to Stockfish; Komodo lost to Jonny in the so-called world computer championship in the same way: it had a positional draw but got attracted to some utterly non-simple option that was microscopically better in its evaluation. The refutation took some time (and was beautiful), but it was there and Jonny went on to win. The important thing to realize is that this wasn't "bad luck". Komodo, like Stockfish in the ongoing TCEC, went from a simple draw - a draw requiring only maintenance - to one (not really, but if it had worked) where the game remained a game and thus more work would be needed to save the game.
I say this not to criticize any programmers. This sort of problem has been around forever, and perhaps it's simply impossible or too costly (from an efficiency standpoint) to fix, at least at this point in time. (Maybe those of you more familiar with the nuts and bolts of programming chess-playing programs can weigh in on how, whether and when such problems can be fixed.) It's interesting that while so many of the things that chess computers supposedly can't do are now routine, this problem continues without any signs of abating.