Taming AI Agents with Harness Engineering

The author moved away from prompt-only control after an agent kept violating forbidden instructions, and instead focused on governing the execution environment.

A 6-step thinking model was simplified into a 4-step workflow, with stack-specific rules split into modules and deployment handled through a Claude Code plugin.

The resulting Smith/Anvil architecture combines STATE and ACT to build agents, then generates static agents at build time rather than interpreting them at runtime.

The case suggests that AI agent quality and safety depend more on harness design, modularity, and deployment structure than on prompts alone.