Thoughts on starting new projects with LLM agents
a day ago
- The project watgo is a from-scratch Go project built with significant help from AI agents, contrasting with a previous Python rewrite.
- Design began with a Markdown file, then agents wrote small, reviewable CLs (changelists); separate CLs for modifications and refactoring were used to keep changes manageable.
- Projects are split into low-importance (vibe-coding) and high-importance (must review all agent code); watgo falls into the latter category requiring full human understanding.
- A practical workflow uses a CLI agent locally with VSCode diff view for review and manual tweaks before committing.
- Small CLs are imperative to maintain human comprehension; agents tend to take shortcuts, requiring guidance toward the big picture.
- A solid test suite is crucial for agent success; for watgo, existing WASM spec and wabt test suites were adapted.
- Go is an excellent language for agent-written projects due to its readability, infrequent changes, minimal idioms, rich standard library, and explicit error handling.
- For learning new subjects, agents should not replace hands-on experience; junior engineers should be cautious, while senior engineers can benefit from agents as productivity boosters.