NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 25·5 min read
Energy per Successful Goal: Goal-Level Energy Accounting for Agentic AI Systems
TL;DR
Current energy benchmarks measure AI consumption per inference, which misrepresents costs for agentic systems (AI agents that orchestrate multi-step workflows with retries and tool calls). Panigrahy and Tyagi built A-LEMS (Agentic LLM Energy Measurement System), which measures Energy per Successful Goal (EpG)—total energy across all execution attempts normalized by completed goals.
✦ Why It Matters
Engineers building agentic AI systems can now measure true energy costs of orchestration decisions, enabling optimization beyond inference efficiency.
Key Takeaways
- Measure agentic AI energy at goal completion level, not per-inference, to capture true workflow costs.
- Use A-LEMS framework to aggregate energy across retries, failures, and multi-step orchestration.
- Orchestration structure, not inference compute, drives 4.33x energy overhead in reasoning tasks.
- For tool-augmented tasks, agentic execution can be cheaper than linear—validate with OOI metric before optimizing.
Full Summary
Current AI energy metrics measure consumption at the granularity of single model invocations, which works for simple single-turn tasks but breaks down for agentic systems—AI agents that autonomously execute multi-step workflows involving tool calls, retries, and failure-recovery cycles. The number of inference calls becomes an implementation detail rather than a meaningful unit of work.
Panigrahy and Tyagi introduced A-LEMS, a cross-layer measurement framework that redefines energy accounting from energy-per-inference to Energy per Successful Goal (EpG), aggregating total workflow energy across all execution attempts and normalizing by successfully completed goals. A-LEMS uses a temporal boundary model, a five-layer observation pipeline mapping RAPL signals (hardware power measurements) to workflow-level energy, and a reproducibility protocol binding measurements to hardware and runtime configuration.
Across eight task families, agentic workflows consumed 4.33x higher mean energy per successful goal than linear baselines. Critically, for tool-augmented tasks, the Orchestration Overhead Index (OOI) inverted below 1.0x, showing agentic execution was cheaper—confirming the metric captures orchestration structure rather than systematic bias.
These findings establish that energy-per-inference is insufficient for agentic AI systems.