A student developer spent four months building a project using AI agents, resulting in a codebase that was rewritten multiple times. The process involved consuming all available free token limits and led to a loss of understanding of the code's internal logic as the AI generated over 1,000 lines every 30 minutes.
The developer tested various AI coding environments, including Codex, Claude Code, Vibe, Gemini-cli, Qwen Coder, Zcode, Hermes, Goose, and Manus. Each tool offers different levels of integration, model support, and virtual environment stability. Some platforms, like Manus, provide isolated containers for agents, while others focus on minimalism.
The author ultimately selected Pi for its modular architecture, which allows users to build custom agent configurations. Pi provides a minimal base of four tools—write, read, edit, and bash—and supports integration with various model providers, contrasting with proprietary solutions that often restrict users to specific models.
The project originated from a university assignment to create an AI agent for a student database and document search. After the initial team disbanded, the author continued development independently to build a portfolio project, exploring the capabilities of autonomous coding agents and their impact on development speed.
