The Harness Margin Opportunity

A UC Berkeley study shows that the software layer that controls AI agents - the harness - largely determines the cost of producing the same output. Across 42 within-model harness comparisons, identical models produced the same results while harness choice drove cost differences as large as 71% (one example: a model charging $1.54 per task on one harness and $0.441 on another). Statistical tests found no consistent quality differences across harnesses, implying that smarter orchestration and routing to cheaper models can deliver identical answers far more cheaply. Harnesses compress repeatable workflows into deterministic code or skills, reserving expensive model calls for the few steps that need them.

That cost delta creates a durable business advantage. In a concrete illustration, two vendors bid on a $250k engagement: one calls a state-of-the-art model at every step and spends $131k on inference (38% gross margin, 19-month sales payback), while a harness-optimized rival spends $37k (75% margin, 10-month payback). The cheaper operator can grow and hire faster and is harder to copy because routing safely requires thousands of repeat evaluations. Building a winning harness requires deep customer understanding, relevant evals, and an automated hill-climbing process. The study notes limits: rollout sorting broke pairing and the sample size only detects large quality swings, so small quality gaps remain undemonstrated rather than ruled out.