The longest circuit ever performed is probably one used for benchmarking. So the point would be to run many gates, look at the associated decay in fidelity, and use that to find a measure of how good the gates are.
Here's a recent paper that I found on a fairly brief search. In Fig. 2 we see depths of 1500 single qubit gates and 130 for gates on two qubits. Given that these circuits are designed to drive states to a point of complete randomness, there's not much point in going further.
However, note that circuit depth isn't uniquely defined, especially since one circuit can be recompiled to an equivalent one with non-equal depth. This is especially true of single qubit operations. For example, should the gate sequence $H Z H$ be regarded as depth 3 or as depth 1, because it is equivalent to $X$. So a more concrete measure is perhaps the depth of cnots.
For this, the longest I know of is my own benchmarking program that I ran for 10 'rounds' in which each round has a cnot depth of 4 (except for the last, which has two). This makes a total cnot depth of 38. But, of course, I only know of this because it is my own project. There might be runs with more cnots, though current noise levels mean that there is not much reason to go too far beyond this.