AI Coding: A Sober Review (www.ubicloud.com)

Over 6–7 months the author ran a pragmatic, experience-based evaluation of AI coding tools used in day-to-day engineering: Continue.dev, Cursor, Windsurf, Claude Code, Cline and various hosted and self‑hosted LLMs (including a self‑hosted Qwen 2.5 on an RTX 3090 and cloud models via OpenRouter from OpenAI, Anthropic and DeepSeek). The findings are subjective but consistent: editor-integrated assistants (Windsurf, Cursor, Continue.dev) are already highly useful for tests, prototyping and repetitive work, while agentic systems like Claude Code and Cline can autonomously run the edit → build → test loop for longer, more complex tasks. Practical wins included Claude Code producing a working fuzz-test scaffold from an OpenAPI spec in ~30 minutes for ~$3 of token cost; Continue.dev recently added deterministic diff apply, fast apply and improved agent mode; Windsurf excels at team workflows and context sharing. For the AI/ML community these observations matter because they underscore where capability gaps and economic tradeoffs still lie: long context windows, persistence and smart scoping materially improve productivity, cloud APIs currently beat most self‑hosting in price‑performance, and top coding models remain largely closed‑source even as open models rapidly close the quality gap. The broader implication is that tooling will keep improving quickly — agentic, context‑aware assistants are becoming viable collaborators, but human oversight remains essential for complex debugging and design.