Key Takeaways:

• The locally deployed open-source model Qwen2.5-72B outperforms GPT-4o with Cline.
• A more general instruction-tuned model is preferable over coding/math-specific versions when using Cline.
• Improvements in model performance could enhance other tools, such as OpenHands, which I plan to try next.

Exploring the Coding Agent Landscape

Recently, one of the hottest topics has been coding agents. Our team explored the Cursor and Windsurf editors (both forks of VS Code) and found them to be amazing! Their capabilities, including code quality, response speed, automation, and less human feedback, are superior to the GitHub Copilot extension on VS Code. It’s not surprising that Cursor and Windsurf can easily outperform GitHub Copilot (and Cline + GPT-4o which I’ll introduce soon), as these products have comprehensive optimizations, including post-trained models, customized prompts, RAG, etc. However, they also come with privacy and security risks.

To adopt a coding assistant while meeting compliance requirements, the Cline extension with the GPT-4o service provided by Azure AI is an excellent option. We are grateful for the leadership team’s encouragement and support in utilizing this tool.

My Short Journey with Cline + AOAI GPT-4o

I started using Cline + AOAI GPT-4o to replace Cursor right after the callout...