Multi-agent scaffolding may not be a permanent artifact: it could just be code that models eventually write for themselves. This AI on Air clip comes from RedpointAI, where @OriolVinyalsML, VP of Research at @GoogleDeepMind and co-lead of @GeminiApp, lays out a forward-looking take on how agent systems are designed. ▷ The complex systems built around models today, multi-agent setups, sub-agent delegation, very long-running tasks, are fundamentally a piece of code. In the limit, the model could write that scaffolding on the fly, producing the most token-efficient, highest-quality set of sub-agents for each task, until maybe no fixed system remains at all. ▷ The reasoning paradigm shift is already underway: the question is no longer just how long a model can reason, but how long it should reason given the complexity of the ask. Automatically generating the right scaffold for the right task is likely the next step. ▷ Making long-running agents more reliable cannot rest on scaffolding and prompt-induced generalization alone. The weights themselves need to train on distributions of long-context tasks, much as the 1.5 long-context breakthrough did, until the model catches up with future use cases.