A new tutorial outlines how to build a Universal Agent that can load specialized skills on demand, effectively acting as a project manager for a 10-person team. The agent swaps between roles like technical-writer or marketing-analyst by loading .skill files, and can even author new skills for itself. It's designed to be simple, requiring only a Next.js project and no complex infrastructure.

The architecture uses an in-memory sandbox with a bash-tool for file operations. Skills are stored as read-only files in S3, while a persistent workspace holds user files that survive across conversations. The system loads both at the start of a request, lets the agent work, and saves changes back to disk on completion, enabling continuous, multi-session development.

This approach mirrors a real startup: the agent is the product manager, each skill is a specialist, and you are the CEO. It sidesteps heavy infrastructure like Docker, focusing on file manipulation and tool calls. The guide details core implementation, from skill structure and route handlers to workspace persistence and a system prompt built for Claude's XML-native format.