We compared old and new agent swarms building SQLite from scratch and found that better coordination delivers similar quality at a fraction of the cost.
I wonder if a strict divide into planner and worker agents really is the best design for swarms. In the spirit of stigmergy, swarm intelligence emerges from every agent performing both of these roles according to their current context. I also wonder how much more performance could be gained by updating model weights as the swarm churns, either of every agent on their own or of every agent class. This would be the truer representation of learned behavior than the field journal described in this article, and it would also implement lessons from context drift and ‘LLM sleep’ from earlier this year. Impressive publication, nonetheless.
I wonder if a strict divide into planner and worker agents really is the best design for swarms. In the spirit of stigmergy, swarm intelligence emerges from every agent performing both of these roles according to their current context. I also wonder how much more performance could be gained by updating model weights as the swarm churns, either of every agent on their own or of every agent class. This would be the truer representation of learned behavior than the field journal described in this article, and it would also implement lessons from context drift and ‘LLM sleep’ from earlier this year. Impressive publication, nonetheless.