In a recent blog post, I explored the capabilities of AI agents, contrasting with the popular buzz surrounding agentic coding. While Large Language Models (LLMs) excel at simple coding tasks, AI agents remain unpredictable and costly. However, recent advancements have sparked renewed interest. Following Google’s release of its generative image AI, I developed a Python package, “gemimg,” and found that emerging LLMs on OpenRouter could enhance and refine the code.
Although early experiences with GitHub Copilot and Claude Sonnet 4.5 were underwhelming, further testing with Google’s Nano Banana Pro demonstrated the potential of integrating AI into practical workflows. This inspired me to employ advanced AI for handling complex tasks, resulting in noticeable productivity gains and quality improvements.
The implementation of an AGENTS.md file, which regulates agent behavior, greatly improved results, ensuring consistency and adherence to desired coding standards. Experiments with Claude Opus 4.5 revealed impressive capabilities when properly guided.
Later, prompted by the release of advanced models by Anthropic and OpenAI, I successfully utilized AI agents to generate practical applications and complex projects, notably expanding my skills in Rust, a language known for its speed and reliability. While AI-generated Rust code performed exceptionally, automating code optimization introduced fascinating possibilities for writing faster, original Rust implementations, showcasing AI’s potential to accelerate programming innovation. These experiences encourage a more optimistic view, despite skepticism, suggesting that modern AI agents offer substantial utility if engaged with clear, structured guidance.