AEQ: measuring the architectural efficiency of AI agents

AEQ, the Agent Efficiency Quotient, measures the architectural efficiency of AI agent systems. Not how good the model is, but how much of what you paid for reached the answer.

Accuracy gets benchmarked on every model release. Efficiency almost never does, so two teams running the same model on the same question can pay wildly different amounts for the same result. AEQ makes that difference measurable: business value delivered divided by tokens consumed, with the numerator held constant by an equivalence rubric so the token delta between architectures is architectural waste by construction.

Token prices keep falling, which is often used to argue efficiency does not matter. Three things do not get cheaper: latency compounds across chained steps, instructions buried in noise degrade reliability, and every wasted token displaces real work in a finite context window.

The specification is the document to cite. It is applied to per-seat software economics in The Cost of a Question and to a working system in the EAM case study.