Seatbelts for AI Agents, Harness Engineering: The Claude Code and Terraform Case

To solve the problems of Loss of Context and Positive Self-Assessment that occur when running AI agents over long periods, Harness Engineering is needed. This is a methodology that systematically designs an agent's scope of work, permissions, and state transitions to limit its scope of thinking.

The following strategies are applied for efficient operation.

• English-based Communication: Reduces token consumption to efficiently maintain the context window.
• Terraform MCP (Model Context Protocol): Dynamically references official documentation to prevent hallucination and ensure code consistency.
• State-based Tool Isolation: Separates permissions according to the planning and execution stages to block indiscriminate infrastructure modifications.

In an actual Azure infrastructure build case, safety was secured through Root module creation, infrastructure governance checks, and a Human-in-the-loop (HITL) structure in which a human performs the final review. In addition, a method of switching roles to a 'security officer persona' to have the agent review its own code is also utilized.