The Agent Harness Is the New Moat

LangChain's deepagents-cli jumped from Top 30 to Top 5 on Terminal-Bench 2.0 with a 13.7-point gain - and the underlying model never changed. Stanford's Meta-Harness hit 76.4% on the same benchmark using Claude Opus 4.6. OpenAI's Frontier team shipped over 1 million lines of production code with zero human-written code using just 3-7 engineers. The harness is now the moat, not the model.