What is an AI harness and what is inside it?

An AI harness is the software a model runs inside. The model supplies the reasoning; the harness supplies everything that lets the reasoning touch the real world: the loop that keeps it going, the tools it can call, the memory it carries between steps, and the permissions that keep it in bounds. The common shorthand is that the model is the engine and the harness is the car. As Addy Osmani puts it, a raw model is not an agent; it becomes one once a harness gives it state, tool execution, feedback loops, and enforceable constraints.^[1]

The pieces around the model are what differ from one tool to the next:

• The agent loop that runs reason-act-observe (see AI agents).
• Tools and MCP: the verbs the model can use. The tool definitions live in the harness, not the model (see tool use and MCP).
• Context and memory: what the harness shows the model and what it remembers.
• Permissions and safety: plan mode, approval steps, sandboxes (see rules and guardrails).

This is why benchmark scores, which measure the model, do not tell the whole story. The harness decides how well that model can see your code and act on it. Swap the same model into two different harnesses and you can get two different agents in practice: one that reliably ships a fix and one that keeps re-reading the same file because its context management is worse, one that asks before deleting a file and one that does not because it was never given a permissions layer to ask through. None of that shows up in a model-only benchmark, which is why how the harnesses actually compare is a separate question from which model is strongest this month.

The four pieces above are not independent add-ons; they depend on each other. Tools are useless without a loop to call them from. Memory is useless if nothing in the loop decides what is worth remembering. And permissions only matter once there are tools capable of doing something worth restricting — a harness with no tools needs no permissions layer at all. That’s the same reason the form factors page and the autonomy ladder both track how much a harness is trusted to do, not just what it is technically capable of.