An agent harness is the software layer that sits around an AI agent and controls how it receives input, decides on actions, calls tools, and returns outputs. It is the orchestration and policy layer for the agent, not the model itself. A harness may handle prompt formatting, tool routing, memory, permission checks, retries, and logging.
This layer matters because it turns a model from a text generator into an operational system. If the harness is weak, an attacker can use prompt injection, malicious tool outputs, or poisoned data to make the agent leak information or trigger unsafe actions. Good defenses focus on least privilege, tool allowlists, input validation, output filtering, sandboxing, and observability. In practice, security teams test the harness as much as the model, because the most serious failures often happen in the control logic that connects the agent to real systems.