When it comes to AI-powered coding people seem to have strong opinions which workflows are "correct" or "modern". I am pretty much the opposite... I think whatever fits to your project / your preference / your way of working is an ideal way.
Sometimes I hear phrases such as "stop using AI as glorified auto-complete in your IDE. Start using local agentic workflows in your terminal. 🤖" I don't agree.
Sure, there are projects or parts of the code where an agentic workflow fits perfectly. But for me personally there are also plenty of occassions where I just use the GitHub Copilot autocomplete functionality or even turn off assistance until I am ready to pass it on (write more tests, come up with suggestions, security review, ...).
It all boils down what makes *you* more efficient. If it does not feel right - it is most likely not the ideal approach.
Right now my personal setup consists of:
- Heavy Refactoring & Architecture → Claude / GitHub Copilot
For multi-file structural refactoring, API interface design, or complex logic where deep reasoning and high context retention are non-negotiable. - Local Agent Execution → Pi + Local/On-Premise Qwen / Gemma
For rapid execution, terminal tasks, and offline coding, I run small, fine-tuned open models (like Qwen 3.8 or Gemma 4) locally using tooling like Pi or previously OpenCode (following the "weak model with great harness is better than great model with weak harness" approach).
This setup gives me the desired reliability (local fallbacks) together with a lot of flexibility. For smaller things I don't need to burn through API credits.
The goal isn't to let AI write unverified code - it's to offload the mechanical plumbing so I can focus on system architecture.