An AI agent coding skeptic tries AI agent coding, in excessive detail
6 hours ago
- #Rust Programming
- #AI Agents
- #Machine Learning
- The author contrasts their experience with AI agent coding against the hype, noting initial skepticism but openness to improvement.
- Experiments with LLMs for coding tasks, like the Python package 'gemimg', showed unexpected utility in code quality improvements.
- GitHub Copilot's Claude Sonnet 4.5 was initially unimpressive for data science work but showed promise with specific tasks like updating 'gemimg' for Nano Banana Pro.
- The introduction of Claude Opus 4.5 marked a significant improvement in agent capabilities, especially with the use of AGENTS.md files for better control over code quality.
- Successful projects with Opus 4.5 include a YouTube metadata scraper, a FastAPI webapp for video analytics, and various Rust projects like 'icon-to-image' and 'miditui'.
- The author explores the potential of Rust for high-performance computing, leveraging LLMs to bridge knowledge gaps and optimize code, achieving significant speedups over traditional Python libraries.
- A novel approach to optimizing machine learning algorithms in Rust, such as UMAP and HDBSCAN, resulted in performance gains over existing implementations.
- The development of 'rustlearn', a Rust crate aiming to replicate and enhance scikit-learn's functionality, demonstrates the practical impact of agentic coding.
- Reflections on the implications of agentic coding for professional development and the open-source community, highlighting both the potential and the skepticism surrounding AI-generated code.