Agentic AI Systems Should Be Designed as Marginal Token Allocators

A position paper frames agentic AI as a single token-allocation economy across routing, planning, serving, and training.

Each layer should optimize marginal benefit against marginal cost, latency, and risk rather than just maximize output quality.

The paper highlights common failure modes such as over-routing, under-verification, and congestion in serving stacks.

This lens could improve compute efficiency and inform better evaluation, training, and risk-aware RL methods.