How dooes multiprocess and the environment behave when using multirun?

[mainrs]

I want to do a sweep over my code. The code does use multiprocessing internally to spawn different agents communicating over gRPC. Once the simulation is done, the main process waits for all the agents via join and then exits.

The main process is annotated using hydra's hydra.main decorator. I am now observing the issue that the main process does not seem to "reset" properly. It is as if the second hyperparameter sweep uses the very same main process.

I can tell that this is the case because internal state stays the same between these runs, even though they should be reset/fresh. Is this normal?

@hydra.main(config_name=..., config_path=...)
def main(cfg):
    # Spawn agents....
    agents = ...

    # Do work.
    
    # Wait.
    for agent in agents:
        agent.join()