Thoughts on starting new projects with LLM agents (eli.thegreenplace.net)

A recent post explores the development of a new project called "watgo," utilizing Large Language Model (LLM) agents to assist in coding and design. Unlike the author's previous successful project rewrite in Python, watgo was built from scratch in Go, highlighting key differences in workflow and agent interactions. The author emphasizes the importance of maintaining human oversight in high-importance projects, insisting on code reviews and iterative refinements rather than relying solely on agent-driven coding, which can sometimes lead to convoluted results due to LLMs generating large changelists (CLs) that require extensive cleanup. The significance of this experience for the AI/ML community lies in the practical lessons learned regarding the integration of LLMs into software development workflows. Key takeaways include the necessity of using a solid test suite to validate agent-generated code and the benefits of employing Go for LLM projects due to its readability and consistency. The author warns junior engineers against overly depending on LLMs for learning, asserting that experiential understanding is irreplaceable. Conversely, senior engineers can leverage LLMs effectively to boost productivity, provided they continue to actively engage with and oversee the code being produced.