Routing at the prompt level seems naive. How is that better than a janky classifier.
What I want is a deeply introspective model that knows to take a step back and call in for aid.
> a deeply introspective model that knows to take a step back and call in for aid
This would mean training/RL a model to know it's getting nowhere and call for help.
But afaik, the only way a model would be able to do that would be review of its own historical output.
And if that's what's being reviewed anyway, doesn't it make more sense to just layer a separate specialized "anti-loop/stuck" model on top to judge the output trends?
It's being done at the harness level. It's still a model making the call.