Chapter 1: What even IS this thing?
Strip away everything else, and a language model just generates text. It can't read files. It can't run code. It can't remember what happened 10 minutes ago. By itself, it's brilliant but powerless.
An agent is a model plus everything else. That "everything else" is the harness. The tools. The loop. The context management. The permissions. The memory. The error handling.
The LLM is the CPU: raw processing power. The harness is the Operating System: it manages resources, provides drivers, handles the boot sequence. Your agent is the application running on top.
- Phil Schmid, "The Importance of Agent Harness in 2026"
Endor Labs tested the same model, same week, in two different harnesses. GPT-5.5 in Codex scored 61.5%. The same model in Cursor's harness: 87.2%.
A 25-point swing. Not from a model upgrade. From switching the wrapper.
Every production harness needs: an orchestration loop, tool registry, permission system, context management, state persistence, error recovery, guardrails, observability, and human-in-the-loop.
Click any component in the diagram to see what breaks without it.
Not in theory. You'll drag components, wire them together, configure them, and run them. Break things. Fix things. The harness is yours to build.
Skip to Sandbox →