at perplexity we are really excited about multi-agent collaboration! we recently explored a simple instantiation of this with advisor escalation from a local model to a remote frontier model, where we showed it can significantly boost the local model's performance. an alternative way i like to think about this: the local model becomes a **gateway** to remote models. its main job is to preprocess all the raw tokens, package them into fewer, more information-dense tokens, and send only those to the server. this preprocessing has several advantages: it reduces cost and latency, enables personalization, and allows privacy controls. already such approaches can save 50% of tokens, but imagine a future where 90%+ of tokens are processed locally and only 10% are sent to the server for the last-mile hardest reasoning. that is a 10x cost reduction and a much more pleasant UX due to low latency. training such a multi-agent system is not an easy problem and still requires a lot of research in multi-agent RL, but we are making good progress and actively hiring for this. if this sounds interesting please DM me