coding-agent benchmarks

What would a fair benchmark for agent architecture look like? [D]

A Reddit user proposes a controlled experimental design to isolate and benchmark architectural components of AI coding agents—specifically workflow decomposition and model routing policies—separating them from model capability to enable falsifiable, component-level evaluation.