Agent = Model + Harness

An agent is defined by the equation Agent = Model + Harness, where the model provides judgment and the harness provides execution capabilities like tool brokering and state persistence. Research on OpenVLHarness demonstrates that a medium-effort harness outperforms maximum-effort setups, including bare models, Visual Sketchpad, and Codex-style code-as-action. This indicates that improving the runtime environment is more effective for real-world visual agentic tool-use than simply increasing reasoning tokens or model effort to solve complex tasks.