Scripted Headless Generation versus Live Agent-Driven Modelling of a Low-Poly Video Game Character: A Single-Case Study

One language model agent built the same six-legged game creature twice from one specification, a 2,800-triangle budget and one acceptance gate: once as a seeded headless Blender script (route A), once by driving an open Blender window through a Model Context Protocol server (route B). Route A's body mesh had 138 sliver triangles against 374, an area coefficient of variation of 0.80 against 1.64 and a vertex cache miss ratio of 1.59 against 2.27, and it kept all 24 teeth of the specification against 22. Route B used half its texture against 28 % and processed 4 % fewer GPU vertices. Route B reached its first passing file in about 8 minutes against 38 and was cheaper in its build phase (USD 2.78 against 3.41 at list price), but once each route's own fixes are counted the order reverses: USD 4.51 for A against 5.44 for B. The reviewer caught four faults live in route B and none in route A. With one run per route, these are measured observations and three hypotheses for a controlled follow-up, not a ranking of the two routes.