Agent Architecture Audit provides a full-stack diagnostic workflow for agent and LLM applications to uncover wrapper regression, memory contamination, tool discipline failures, and hidden repair loops. It walks through a 12-layer evaluation stack to inspect system prompts, session history, memory admission, and rendering pipelines. It produces severity-ranked findings with code-first fixes instead of generic prompt tweaks. Use this skill when releasing an agent application to production, when agent behavior inexplicably degrades, or when debugging complex multi-step workflows.