This AI Agent Coded My Entire App Unsupervised (What Could Go Wrong?)

A hands-on walkthrough of setting up OpenClaw, an AI personal assistant that controls your computer, to autonomously build an iOS app via Telegram messages. The author runs OpenClaw inside a UTM virtual machine on macOS for safety, installs Xcode and Claude Code, and instructs the agent to build a SwiftUI to-do app remotely. Key concerns covered include prompt injection vulnerabilities, the risks of giving an AI agent full computer access, and how MCP (Model Context Protocol) via Zapier can limit agent permissions to specific tools. The experiment succeeds in producing a basic working app, but setup is complex and the agent requires monitoring to avoid loops and runaway API costs. Practical safety advice: use a VM or spare machine, restrict permissions via MCP, and avoid leaving the agent unsupervised.