I've been playing around with agent-native source annotation to specifically address the massively parallel work problem. Check it out here: https://github.com/draxl-org/draxl