LLM Harnesses: Why the Wrapper Matters More Than the Model

The article argues that an LLM agent’s success depends more on the surrounding harness (tools, memory, context management, and verification checks) than on model selection. It cites Vercel’s v0 team removing 80% of their agent’s tools to lift success from 80% to 100%, and Claude Opus 4.5 scoring 42.22% under one scaffold versus 77.78% under Claude Code on the same benchmark, and includes a runnable JavaScript harness example using the Anthropic SDK.