Aravind Srinivas reframes what a “computer” is. The next evolution of computing isn’t a better model — it’s the orchestration system that conducts many models into a unified intelligence. Perplexity launched a massively multimodel product: 19 backend models functioning as a single agent system, capable of delegating tasks, managing files, executing code, and browsing the web within a secure sandbox.
“No one model family can do its best work for you without the talents of others.” Just as a conductor arranges musicians’ talents, AI systems must coordinate specialized models. The differentiation isn’t individual models but “how you orchestrate them.” This extends Multi-model code review creates adversarial robustness — each model catches what others miss from code review to the entire product: models with different strengths and failure modes create emergent capability that none achieves alone.
Srinivas argues this fundamentally changes how we interact with information: “when AI can find, analyze, reason, and accurately present all knowledge on the open web,” the way we consume the internet changes. The orchestration layer is what makes this possible — no single model can find, analyze, reason, AND present. Different models handle different parts of the pipeline.
Srinivas traces computing through multiple eras, each returning more autonomy to machines and freeing humans from watching interfaces. AI as orchestration system is the next step — the computer handles multi-step reasoning autonomously. This is An orchestrator agent that manages other agents solves the parallel coordination problem without human bottleneck at the infrastructure level: the meta-agent isn’t just managing sub-agents, it’s managing sub-models as a unified system.
With new frontier models arriving rapidly, the symphony expands continuously. This is why Production agents route routine cases through decision trees, reserving humans for complexity matters more than ever — as the model zoo grows, deterministic routing becomes core infrastructure. And it vindicates Treat AI like a distributed team, not a single assistant: 19 models working as a team outperform any single genius model working alone.