Costs of longer agentic task lengths

A blog post builds a turn-based Markov model of agent-assisted coding, in which longer autonomous turns raise the risk of wrong abstractions that must later be rolled back. With its default parameters, a taste-0.2 engineer peaks at 2.04× an expert's pace with 1.2-hour turns, and a taste-1 engineer at 3.11× with 3.2-hour turns. Quadrupling the time horizon adds only about 0.3×, and two parallel agents is optimal. These figures come from the author's own assumptions, not measurements, so they show a dynamic rather than a forecast.