When AI work runs in a person's chat, the question of who runs it answers itself. Once work runs on a schedule or is handed off, it needs an answer written down: which agent, which version, which model, with which credentials, on which machine, for how long. An executor is that answer, given a name so a process can refer to it instead of repeating the details.
Executors separate the process from the runtime. The workflow says what must happen at each step; the executor says who does it. That lets the same workflow run attended today and unattended later, or let a cheap model handle routine steps while a stronger one handles the hard step, without rewriting the process.
The reliability questions live here too. A good executor setup has a time limit per run, a way to stop a run, a record of what each run cost, and credentials handed to the run that needs them and to nothing else.