Researchers are addressing fundamental infrastructure challenges for deploying AI agents across edge and cloud environments, according to four papers published on arXiv.
A paper titled “Where Should Agents Live?” introduces agentic-eCAL, a metric for measuring energy costs in multi-agent workflows. According to the research, which analyzed hundreds of GPU benchmark configurations on NVIDIA A100 and H100 across 16 open-weight models and 8 orchestration topologies, inter-agent text transport incurs only 0.25% of workflow energy across 5G RAN, metro, and optical links. The findings demonstrate that “the dominant energy cost of distribution is often not the transmission of inter-agent text itself, but the additional inference and context processing induced by that communication,” according to arxiv.org.
In a separate paper on “Collaborative Memory for Multi-Agent VLM Systems,” researchers propose a framework for vision-language model agents to share visual context. According to arxiv.org, the framework addresses how “agents to build on contributions from other agents, recover missing visual context, and reconcile differing interpretations as new evidence emerges.”
Another paper introduces XConf (eXperiential Confidence), a confidence estimation method that stores “a record of the model’s own graded past episodes.” According to the research, XConf beats or matches ten-sample self-consistency in discrimination on 23 of 24 comparisons across nine benchmarks, at a tenth of the generation cost.
Finally, rMuscle, a real-time VLA inference framework for robotics, achieves 1.29-1.42X speedup on RTX 4090 and Jetson Thor by exploiting “cross-execution similarity through a dual-phase muscle-memory cache,” according to arxiv.org.