A Tale of Three Months with AI: Guardrails, and How We Are More Alike Than We (alessandrovermeulen.me)

In a recent reflection on the evolving landscape of using AI for software development, an engineer shares their journey over three months experimenting with large language models (LLMs) for agentic development. Initially skeptical about the capabilities of models like Codex and Gemini, the engineer grew enthusiastic as they utilized these tools to tackle complex tasks, such as firmware inspection and cloud infrastructure provisioning, illustrating how LLMs can effectively simplify the coding process when tasked with clear, specific objectives. However, as project complexity increased, the limitations of LLMs became apparent, particularly their declining performance with expansive codebases and the need for more explicit instructions. This experience signifies a deeper understanding of how AI models function similarly to humans, responsive primarily to constrained tasks rather than open-ended inquiries. The engineer advocates for a structured workflow that integrates rigorous checks and balances, emphasizing the critical role of maintaining architectural clarity to avoid the pitfalls of project decay. The takeaway is that while LLMs can enhance productivity in greenfield projects, proper oversight and decision-tracking are essential as complexity grows, suggesting that a disciplined approach could cultivate more manageable, productive development environments in AI-driven coding.