agent swarms (like single agents) are myopic in their problem solving, prone to local minima & rabbit holing
two swarm design methods to get around this are
1. set up your problem as an adversarial game that multiple swarms compete on (requires good Goal/Task design)
2. explicitly seed multiple swarms with different ideas and let them work independently and cross pollinate (fork & flush approaches)
ex: been experimenting with swarms to drive an eval factory
Task proposer -> hacker -> external judge
the intended design is that the goals of the swarms play against each other without collapsing (tbt GANs) & the shared knowledge improves the entire system