UC Berkeley Study Reveals Coding Agent Framework Overhead Can Double Costs Without Improving Accuracy

Melissa Pan and researchers at UC Berkeley's Sky Computing Lab have published the first systematic study quantifying the "harness tax"—the hidden token cost imposed by coding agent frameworks—revealing that different harnesses for identical AI models can vary costs by up to 5× while maintaining nearly identical success rates. The research, published on the HarnessTax project site, tested 21 model-harness combinations across benchmarks including SWE-bench Lite and Terminal-Bench 2.0, examining how the software layer surrounding an AI model affects both expense and performance.