Harness-Delta Attribution Dissects Gains in Agentic Systems
October 19, 2026
Harness-Delta Attribution (HDA) is a new protocol to determine if performance gains in optimized agentic harnesses come from actual generalization or benchmark-specific overfitting and test-time scaling. Testing across 11 models showed that a significant portion of gains in many settings are attributable to overfitting.
HOW THIS AFFECTS YOU
●
builderThis warns you that improvements in agentic performance may be due to prompt optimization rather than model intelligence.
●
researcherYou can use HDA to verify if your agent orchestration improvements actually generalize.