agentic-eCAL puts an energy number on where a multi-agent team should physically run

arXiv·low signalXBlueskyLinkedInCopy link

arXiv 2609.18283 (16 Sep 2026) generalizes the Energy Cost of AI Lifecycle metric from single-model inference to directed multi-agent workflows, coupling a closed-form two-rate model (compute-bound prefill, memory-bound decode) with 7-layer OSI data transport so inter-agent communication is priced rather than ignored. It is grounded in hundreds of GPU benchmark configurations on A100 and H100 across 16 open-weight models and 8 orchestration topologies. The framing is telecom 5G-Advanced and 6G operations, but the question it answers — does agent-to-agent chatter cost meaningful energy, and which tier should hold the team — applies to any edge-cloud agent deployment.