arXiv cs.AIMonday · May 25, 2026FREE

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

A new research paper from arXiv introduces A-LEMS (Agentic LLM Energy Measurement System), a cross-layer measurement framework designed to address the limitations of current AI energy benchmarks for agentic AI systems. Traditional benchmarks measure energy consumption at the granularity of a single model invocation or training run, which is coherent for single-turn workloads. However, for agentic systems that involve multi-step orchestration, tool calls, retries, and failure-recovery cycles for a single user goal, the invocation count becomes an implementation artifact rather than a true task property, misrepresenting the actual energy cost of goal completion. A-LEMS redefines the unit of AI energy accounting to Energy per Successful Goal (EpG). EpG aggregates the total workflow energy across all execution attempts, including those that fail and require retries, normalizing this sum by the number of successfully completed goals. The framework formalizes energy attribution through a temporal boundary model, a five-layer observation pipeline that maps RAPL signals to workflow-level energy, and a reproducibility protocol linking measurements to hardware and runtime configurations. Building on EpG, the paper also defines the Orchestration Overhead Index (OOI), which isolates the energy cost specific to orchestration.