Agents allow compartmentalisation of capabilities and access, which I think can open many use cases not possible with single models

And that’s to say nothing about the preservation, or rather distribution of context, between agents, potentially allowing much better performance on long task endeavours