The author evaluated AI tools for software development, including Cursor with Claude, GPT-4o, GigaCode, and local models like Qwen2.5-Coder and DeepSeek-Coder. The workflow was restructured to prioritize intent formulation, specification, and risk assessment, which reduced planning time to 30–60 minutes.
AI agents demonstrated a speed increase of 2.5–4 times for routine tasks like CRUD operations and API integrations. For complex business logic and architectural decisions, the speedup was 1.3–1.8 times. The author noted that AI often generates insecure code, including SQL injections and hard-coded secrets.
To mitigate risks, the author implemented a security checklist, manual reviews, and property-based testing. While AI handles unit tests well, complex scenarios still require significant human oversight. The author emphasizes that the primary role of a developer is shifting toward orchestrating agents and ensuring quality.
The author concludes that measuring success based on code generation speed is misleading. Instead, the focus should be on the time required to reach a secure, maintainable, and functional result. Open-source models are effective for daily tasks, but closed models currently offer superior quality for complex projects.
