Very much agreed. I think this could be exacerbated by the fact that many deployment prompts are themselves optimized (DSPy, prompt ablations, etc.) — and optimized prompts tend to be weird. This is prompts for now, but plausibly agent scaffolds and multi-agent configurations soon too. Very ood from the training process.
Very much agreed. I think this could be exacerbated by the fact that many deployment prompts are themselves optimized (DSPy, prompt ablations, etc.) — and optimized prompts tend to be weird. This is prompts for now, but plausibly agent scaffolds and multi-agent configurations soon too. Very ood from the training process.