Model-driven model selection "as a tool"

Feature Requests
1

Feature request for product/service

Cursor IDE

Describe the request

All the major models currently available are around 95% reliable. However, the remaining 5% is not the same between models. That is, when Claude Sonnet 4.5 gets stuck in a confused loop, often switching manually to Gemini 2.5 Pro or GPT 5 will solve the problem. But this could be automated!

I propose allowing model selection be a tool available to the agent. The full solution would be an orchestration agent, and then a swarm of smaller model-specific agents. But the short-term solution is simply to allow (e.g.) Claude Sonnet 4.5 to realise it cannot solve the problem, and switch itself to Gemini, perhaps with a summary of the problem and a “request for help from a fellow model”. A bit of scaffolding could include:

  • Switch back to the original model as soon as possible (the next model makes the tool call)
  • Switch back to the original model after N turns automatically (again, perhaps with a brief description of what happened with the other model)

This would be a huge quality of life improvement, as I think many people are manually switching models when they get stuck.

2

By the way, I believe this is very different from the “auto” model selection. This approach still retains transparency for users and allows them to change the model manually on-the-fly.