Do not accept "magical" behaviour.
I came across quite some projects which had Cypress E2E tests which were, according to the developers, "unstable".
So what is the magic part which makes it fail? There is no such thing as an instable test as such in a test environment, in this case it was a broken test with unmocked data which occasionally threw errors, depending on the OS and speed of the machine it was running on.
In another scenario the was a random number involved, which randomly chose a picture in the test development and depending wether it needed cropping or not, an additional popup would open.
I was told: "well, sometimes it works, sometimes it doesn't". Nobody actually investigated further, tests were excluded instead. Noone ever looked at the artifacts/screenshots/logs.
Always investigate, there is a reason for every "magical" behaviour.