Claude Code documentation defines harness as layer around model

The article argues that differences in AI output between users are often due to variations in their harnesses rather than the model's intelligence. It identifies three common issues—varying results, repetitive explanations, and safety concerns—as stemming from missing or inconsistent harness components. The author shares their company's specific setup, which includes one company-wide instruction file, eight department-specific files, 57 prohibited items in settings, five skills, seven sub-agents, and two scheduled tasks. These components were added incrementally as problems arose, such as creating a skill for image insertion after initial failures.

The text distinguishes the harness from the model, noting that while the model thinks and generates text, the harness prepares tools and materials. It also mentions the Claude Agent SDK, which allows developers to use similar mechanisms in their own programs. The article serves as the fifth installment in a series on AI agent terminology, aiming to help beginners understand why the same AI can produce different results depending on the surrounding configuration.