The software layer wrapped around a model that turns it into an agent: tools, a system prompt, context-window management, and permissions. The model alone just takes text in and predicts text out — it can't read a file, run a command, or remember the last turn. The harness supplies all of that: it assembles the context for each request, executes the tool calls the model asks for, feeds results back in, and decides when to ask you for permission. Claude Code is a harness. Cursor is a harness. Claude.ai is a harness too — just a chat one instead of a coding one.
Why it matters: same model, different harness, different behavior — Claude Code edits your files and Claude.ai just answers questions, even on identical prompts. When a product behaves oddly, the fix is often in the harness (system prompt, tools, permissions), not the model. Process discipline like checklists, phase gates, and checkpoints is what you build on top of a harness — see orchestration — not the harness itself.