New A-LEMS metric reveals agentic AI uses 4.33x more energy per goal

Current AI energy benchmarks measure consumption at the inference or training level—fine for single-turn models but misleading for agentic systems that chain multiple steps, tool calls, retries, and failure recoveries. A new arXiv paper introduces A-LEMS (Agentic LLM Energy Measurement System) and a fundamental shift in accounting: Energy per Successful Goal (EpG). EpG aggregates total workflow energy across all attempts (including failures) normalized by completed goals. The framework formalizes energy attribution with a temporal boundary model and a five-layer observation pipeline mapping RAPL signals to workflow-level energy. It also introduces the Orchestration Overhead Index (OOI) to isolate orchestration energy from pure inference cost.

Across five reasoning and three tool-augmented task families, the results are stark: agentic workflows consume 4.33x higher mean EpG than linear baselines (888.1 J vs 205.3 J). Critically, this overhead is driven by orchestration structure, not inference compute. For tool-augmented tasks, OOI actually drops below 1.0x—agentic execution can be cheaper than linear, confirming the metric captures true orchestration efficiency rather than a fixed penalty. The findings establish that energy-per-inference is insufficient for agentic AI, and that EpG and OOI provide the necessary measurement foundation for accurate, comparable benchmarking of autonomous AI systems.