What the harness controls
An AI agent harness is the operational layer around a model. It determines what an agent knows, what it can do, where it runs, and how you evaluate its results.
- Select the context sources the agent needs.
- Grant only the tools required for its task.
- Define actions that require approval or remain blocked.
- Choose a model based on capability, latency, cost, security, and residency.
- Create evaluation cases and acceptance criteria.
- Review run outcomes before expanding access.
Prompt instructions are not a security boundary. Enforce permissions and approval requirements in the runtime and tool layer.
Model-agnostic operation
Keep context, tools, policies, and evaluations outside the model. This lets you compare or replace model providers without rebuilding the operational controls around an agent.
