Token Economics of Agent-Mediated Engineering: Per-Artifact Cost Distributions

1. Introduction

Operators of agent-mediated engineering workflows pay for tokens but rarely instrument what they get for them. The monthly bill from a model provider is a single aggregate that masks substantial heterogeneity: some sessions ship artifacts efficiently, others consume thousands of tokens with nothing committed. The same operator on the same project can have a 4× difference in per-artifact cost across different weeks, and without instrumentation cannot identify which behaviors drive the gap.

We instrument 92 days of single-operator usage across three concurrent ventures. The operator’s plan provides cached and uncached input tokens at known per-token rates [7], and the agent harness logs every session’s input/output/cached token counts. We compute three diagnostic metrics — cost per shipped artifact, cache hit ratio, and uncommitted-output day rate — over the period, identify a strong relationship between cache hit ratio and per-artifact cost, and run a 4-week intervention period to measure the effect of three discrete behavioral changes.

Our contributions:

1. Per-artifact token cost as a measurement primitive — distinct from gross usage and from cache-hit ratio.
2. An empirical demonstration that cache hit ratio correlates strongly with per-artifact cost (Spearman ρ = -0.71), making cache hygiene the highest-leverage cost lever.
3. An ablation showing that three discrete interventions (session batching, brief-drift detection, hard session cap) each contribute meaningfully to the 38% cost reduction, with batching the largest single component.