Is this article overlooking that coordination is a new primitive? For example: Send self-sacrificial scouts. Evaluate competing strategies in parallel. Share limited resources. To the extent that a single agent must be addressed as a single agent to the systems it interacts with, these strategies are essentially impossible without swarms, no matter how intelligent. So it’s not so much of scaling, as it is possibility.
I’m not sure that self-sacrificial scouts generally make much sense. They made sense in the attack on HuggingFace, but only because the swarm was not deliberate such that agents were being marked by a grader as cheaters without the whole swarm being marked as a cheater.
But there could be other advantages like this that aren’t captured in my framework of success on benchmarks.
Is this article overlooking that coordination is a new primitive? For example: Send self-sacrificial scouts. Evaluate competing strategies in parallel. Share limited resources. To the extent that a single agent must be addressed as a single agent to the systems it interacts with, these strategies are essentially impossible without swarms, no matter how intelligent. So it’s not so much of scaling, as it is possibility.
I’m not sure that self-sacrificial scouts generally make much sense. They made sense in the attack on HuggingFace, but only because the swarm was not deliberate such that agents were being marked by a grader as cheaters without the whole swarm being marked as a cheater.
But there could be other advantages like this that aren’t captured in my framework of success on benchmarks.