The one place where I think it works well is implementing a well defined protocol. If you have exact definitions for inputs and outputs, then writing the tests first can make things a lot easier. But it's rare when you have such a clean problem to solve.
As an example, any data infrastructure software that uses non-trivial adapative scheduling to optimize concurrency and throughput famously has this property. It is also the primary reason things like database engines are effectively untestable until the code is nearly complete. In principle you could architect any software to be amenable to TDD, but for these cases the performance penalty implied by that architecture is so severe (literally on the order of 10x) that no one serious would design their software that way.
In these cases, the same binary can pass testing on one machine and fail on another. It is a different discipline than classic TDD, and more critical as more software systems are implicitly distributed.