LLM Agent Performance Tied More to Harness Than Model, Paper Argues

A recent position paper published on arXiv (2605.23950) posits that, for long-term tasks assessed using advanced capability models, the agent execution harness—responsible for context creation, tool interaction, orchestration, and verification—often plays a more pivotal role in determining performance than the language model itself. The authors introduce the Binding Constraint Thesis, asserting that in this context, variations in performance are influenced more by the configuration of the harness than by the model selection, with existing evaluation methods incorrectly attributing improvements at the harness level to enhancements in the model. This thesis is backed by a control-theoretic framework that views the harness as the controller within a closed-loop dynamic system, with the LLM acting as the stochastic policy it regulates, clarifying why minor adjustments to the harness can lead to greater performance changes than model replacements.