Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems

In three linesA-LEMS proposes Energy per Successful Goal (EpG) metric for agentic systems instead of per-inference energy. Across 8 task families, agentic workflows consume 4.33x more energy than linear execution (888.1 J vs 205.3 J). Overhead driven by orchestration, not compute.

Summary generated by Claude — human-verified