Clean Architecture for AI Agents: Designing the Engineering Harness

The article from Habr, based on OpenAI's 'Harness engineering' concept, discusses how to design an engineering harness for coding agents. The harness is defined as a repository-local engineering system that turns human intent into a bounded, reproducible, and verified product change. It consists of three layers: orchestration (managing the process), semantic (providing a product model), and validation (ensuring evidence and independent review). The article applies Robert Martin's Clean Architecture principles to this harness, emphasizing that the goal of architecture is to minimize human effort over the system's lifetime. The key metric is the total cost of a reliable change, not just speed or token cost. The harness has its own policy (authority, lifecycle, permissions) and details (models, Git, CI), and dependencies should point inward. The article suggests a Screaming Architecture where the top-level structure reflects the change lifecycle, not the mechanisms. A practical test is whether a detail can be replaced without affecting the use case contract.