The original title is "Model + Harness = Agent: The Gap Isn't Where You Think"

Short summary

The same model can perform noticeably differently depending on its harness—the surrounding system for context selection, tool exposure, permissions, memory, and verification. Moonshot's own K3 benchmarks show harness swaps moving SWE-bench performance by up to 15 percentage points, dwarfing typical model-to-model gains of 2-4 points. The harness, not the model, is often the thinner pillar limiting what agents can actually do.

• •Model and harness are separable layers; harness quality can swing benchmark scores by 15+ points
• •Moonshot officially discloses harness as an evaluation condition, publishing Claude Code integration for K3
• •Harness components: context engineering, tool safety, human interaction, memory, multi-agent orchestration

Generated with AI, which can make mistakes.