The Physics of Agent Work

Newton published his laws of motion in 1687. Force alone still could not tell engineers what a machine delivered. In 1829, Coriolis formalized that quantity as work: effort applied over distance. A joule is a joule: on a steam engine, on a server rack, in a data center in 2026. But energy consumed is not the same as useful work delivered.

Now look at AI. We bill our systems in tokens. A token is not a physical unit, and its energy cost is not fixed. It depends on the model, the hardware, and how the request is served. Two identical token counts can hide very different energy costs. That was tolerable while AI was a component. It stops working with agentic AI. An agent doesn't answer. It works. It loops, retries, calls tools, and burns context until the task is done. Or fails. Tokens measure activity. They don't measure work.

Research groups have started evaluating agents in joules: energy per completed task, not per model call, and value delivered per joule. As agentic AI scales, energy becomes a constraint on what we can economically deploy.

The platforms that understand this will measure their agents the way engineers have always measured machines: not by how hard they push, but by what they move, per joule spent.

Physics holds. Our metrics should too.