Goal-level energy accounting for agentic AI

Agentic systems can hide a lot of cost behind retries, tool calls, and failure recovery. A new paper argues that measuring energy at the single-inference level misses that reality.

The authors propose **Energy per Successful Goal (EpG)**, a workflow-level metric that sums total energy across all attempts and normalizes by completed goals. They also introduce the **Orchestration Overhead Index (OOI)** to separate orchestration cost from linear execution under the same task criteria.

In their experiments across five reasoning task families and three tool-augmented task families, agentic workflows used **4.33x more energy per successful goal** than linear baselines, with mean energy of **888.1 J vs. 205.3 J**. The paper says this overhead is driven by orchestration structure, not just inference compute.

For builders shipping agents, the takeaway is straightforward: if you care about production cost, measure energy at the goal level. Track retries, tool calls, and recovery paths as part of the bill, and compare orchestration designs using a completion-normalized metric.

One nuance: the paper also reports that for tool-augmented tasks, OOI can fall below 1.0x, suggesting agentic execution can be cheaper than linear baselines in some settings.