The AI harness (aka agent scaffolding) is everything around the model that is not the model: the agent loop, tool dispatch, memory, context management, execution environment, feedback, guardrails, and observability.

agent = model + harness. A model alone can only produce text. The harness is what lets it act over multiple steps, use tools, survive its own context window, and run a long task across sessions. Two teams using the same model ship completely different products because of the harness.

📚 Resources