ML//agent//agent harness

The agent harness is the ordinary software that turns a model's generations into a process acting on an environment. An LLM can emit text or a structured function call, but it cannot open a file, retain durable state, or deploy software by itself.

The agent harness is the ordinary software that turns a model's generations into a process acting on an environment. An LLM can emit text or a structured function call, but it cannot open a file, retain durable state, or deploy software by itself.

It assembles context, exposes tools, validates calls, enforces permissions, executes implementations, serializes observations, handles failure, and starts the next inference. The model proposes a transition; the harness decides whether and how that transition becomes real, which makes it the place where agent containment is enforced.

This is wider than a system prompt and narrower than the whole product. A prompt supplies information. An API supplies operations. The harness owns the loop connecting inference, operations, state, and policy enforcement.

The model is the agent's policy, and the harness is its operating system: an agent exists only as a model running inside such a loop.