Agentic coding notes from Galapagos Island
The author discusses their experience using AI coding tools for writing a blog post, analyzing issues with agentic coding loops—such as context loss and tools degrading the codebase—and explaining how specific prompting strategies can improve outcomes. They also cover practical limitations of AI tools in complex codebases and reflect on the tradeoffs of different AI-assisted development workflows.
Background
Dan Luu is a well-known software engineer and writer who frequently posts detailed, empirically grounded critiques of software engineering practices and industry hype. This post is about "agentic coding" — the use of AI agents that can autonomously write, test, and iterate on code across multiple steps (rather than just generating one-off snippets). The title is a joke: "Galapagos Island" is Dan's personal blog. The piece examines how much of the claimed productivity boost from AI coding assistants comes from the agentic loop (the AI running, checking, and fixing its own output) rather than from the model itself. It also warns about "vibes-based" evaluation — judging an AI tool by how impressive its demos feel rather than by rigorous measurement, which leads to misleading claims about capability. Key context: there is currently intense industry debate about whether AI coding tools (Cursor, Copilot, Devin, etc.) truly make engineers dramatically more productive or whether gains are exaggerated, and Luu is known for pushing back against uncritical hype.