#ai-guardrails

Your AI Agent Is Only 1.6% "Intelligence": The Real Battleground Is the Harness Runtime and Guardrails

If you pull an AI agent apart, almost none of it is the model. One analysis of Claude Code found ~98.4% of the code is harness engineering — permissions, context, sandboxing, tool routing, recovery — and only ~1.6% is AI decision logic. That ratio is the whole story of enterprise deployment: agents stall on the way to production because of the runtime and the guardrails around it, not the intelligence inside them. Here is what the harness actually does, why human-in-the-loop is quietly becoming a placebo, what an unbounded agent did to a real booking system, and a five-step FDE checklist for wiring an agent into enterprise systems safely.