Agent Architecture Audit

This is a diagnostic workflow that systematically checks the 12-layer stack of agent applications for common failure patterns like wrapper regression, memory pollution, and hidden repair loops. You run it when your agent worked fine in the playground but breaks in production, or when users report degrading behavior after you've added prompt layers or memory systems. It searches your codebase for anti-patterns (tool requirements only in prompts, not code), maps failures to specific layers, and gives you severity-ranked findings with code-first fixes. The core insight is that most agent failures happen in the wrapper layers between the model and the user, not in the model itself. Worth running before any production release with tool calling or multi-step workflows.

Install to Claude Code

npx -y skills add affaan-m/ecc --skill agent-architecture-audit --agent claude-code