I’ve noticed something interesting while building software with AI agents: sometimes the fastest way to fix a stuck coding agent is simply to switch to another model.
I started building software again using GitHub Copilot CLI. One thing I noticed is how each model is slightly different. The models themselves keep improving and keep getting better; but each one almost has like a personality to it.
Not only the models are improving, but the tools are as well. There are people online who went from Cursor to Claude Code and now over to OpenAI’s Codex tooling. Even OpenClaw’s creator switched from Claude Code over to OpenAI’s Codex recently.
I haven’t gone that far yet, but one thing I have noticed myself doing is switching between Claude Open 4.6 and OpenAI’s gpt-5.3-codex as the coding model I’m using in Copilot CLI.
Why you ask? Because I feel like sometimes the LLMs get stuck on a problem, and putting another LLM in the same context window somehow breaks that loop. My typical process boils down to:
- Start a plan with opus 4.6.
- Start coding with opus 4.6.
- If I notice bad results or the agent getting stuck, switch to gpt-5.3-codex.
I don’t know if I’m the only one doing this, but it has helped me in more than one occasion. Recently I hit this while building a custom agent for Teams. Opus 4.6 got stuck generating the Teams app manifest and kept giving me incorrect manifest files. I switched to GPT-5.3-Codex with the exact same context, and it produced a correct manifest in a single prompt.
I have seen more advanced use cases of this online, where people are having different LLMs do competitive code review. I’m not that advanced yet.
However, I am m interested to see where this goes next. It’d be interesting for tools like GitHub Copilot CLI to automatically select another LLM in case it notices that it’s stuck on a certain bug and keeps looping around.
For now, and for me, one thing stands out: Sometimes the best way to move forward is not to push the same model harder, but to change the mind that is solving the problem.