A research team working with Apex Fusion reproduced the MIT SwarmWorld experiment using an open-weight model on four RTX 3090 GPUs. The study confirmed that AI agents in a shared environment build more resilient technology portfolios than solo agents, even with a compressed model. The replication found that the performance gap between group and individual agents widened with the weaker model, suggesting the collective effect is robust. Additionally, the team discovered that KV cache precision impacted results more than the difference between model sizes. These findings imply that monitoring agent coordination is relevant for anyone running such systems on self-hosted infrastructure.
Read the full story at thenextweb.com.
Photo: nanadua11 / Pixabay
