Simon Willison demonstrated the potential of coding agents by creating a working HTML5 parser while engaging in family activities. His approach relies on a robust testing framework, where unit tests and type checks provide immediate feedback. This method allows the coding agent to operate with minimal supervision, moving forward iteratively and efficiently. Geoffrey Huntley’s Ralph Wiggum loop concept automates this process: an AI agent reads specifications, executes tasks, and verifies results in a clean state for each iteration, preventing context overload. The effectiveness of this approach is illustrated by Willison’s project, which passed over 9,200 tests from the html5lib-tests suite. The complexity of HTML5 parsing is mitigated by the clarity of the tests, which ensure the AI remains grounded in reality. In a broader context, research from Cursor shows that multiple agents can collaborate on a codebase, exemplified by their experiment where they built a web browser from scratch in a week, generating over a million lines of code. While the Ralph Wiggum loop excels in areas with clear verification, it faces challenges in more subjective tasks, like refining web pages to meet brand standards. The article emphasizes the difficulty in defining success criteria, a task often left to human judgment. As AI handles more of the implementation, the role of humans shifts toward setting initial directions and refining outcomes, highlighting the evolving landscape of software development where clarity in objectives enables greater autonomy for AI agents.