Consider an endgame like the one shown below (white to move):
[fen "5k2/5p2/r7/8/6R1/4b1PP/8/5R1K w - - 0 50"]
Stockfish gives white a solid +3 advantage here and that seems appropriate given that white is up the exchange with an extra pawn. I don't have strong endgame technique though, so when I try to play this position against the engine, I can't seem to win. I keep running into various tactics including forced draws, pins, forks, etc, or at least can't figure out how to make progress. If I kept playing against the engine in this position and also used the engine for help when getting stuck, I imagine I'd eventually be able to drive the victory home unassisted by recognizing the various defenses black can employ as well as the resources white has for making progress.
What's unclear to me about this and other endgame positions is the extent to which the difficulty of playing against the engine is due to (1) a lack of endgame technique vs (2) the engine's calculation advantage. In other words, would having solid endgame technique allow you to reliably win endgames such as the one above, or would it often be the case that the engine is able to find refutations to seemingly good moves on the basis of its ability to just churn through possible positions? A human player might avoid certain moves because the resulting positions look like they trap a piece or invite a tactic or allow a forced draw, but an engine of course knows definitively what the possibilities resulting from each move will be (up to some depth).
The possibility that some winning endgames are either difficult or impossible to win against an engine even with well-grounded play seems especially likely (at least from my perspective) for positions involving knights and bishops. At least with rook and pawn endgames, it's easier to calculate given their ability to cut off files and ranks and stick behind pawns. With knights and bishops, however, it seems to require much more accurate play to avoid running into strong coordination sequences from the engine.
One notable example of an endgame containing an outcome-changing move that seems near impossible for a human to find comes from a game between Carlsen and Caruana
[fen "5k2/8/5pK1/5PbP/2Bn4/8/8/8 b - - 0 1"] [StartFlipped "0"] [StartPly "139"] 1... Bh4 2. Bd5 Ne2 3. Bf3 Ng1!!
Black can surprisingly win with 68...Bh4 69. Bd5 Ne2 70. Bf3 Ng1!!, though the sequence seems out of the reach of a human player. Looking at this position on Lichess it seems move isn't even found with the cloud analysis used which is stated to be Stockfish 13+ at depth 72. What I'm curious about is how often and in what conditions this is possible - that a position with a seemingly straightforward outcome (according to GMs) has moves that only an engine could realistically find that lead to a separate outcome.