Perplexity's hybrid agent routes sensitive tasks to local Macs, keeping data on-device

Perplexity has introduced hybrid compute for its agentic platform, Computer, enabling a single AI agent to split tasks between cloud-based frontier models and smaller open-weight models running locally on Apple silicon Macs. The system dynamically hands off confidential portions of a task to the local model without restarting or losing context, ensuring sensitive data never leaves the device. The company claims this is the first such capability for AI agents, with the feature rolling out to users now.

The hybrid compute approach addresses a growing tension in AI agents between the power of large cloud models and the privacy demands of certain tasks. By running open-weight models locally on Apple silicon, the system can process sensitive segments without transmitting them to remote servers.

The seamless handoff between cloud and local models preserves conversational context, meaning users don't need to restart tasks when a confidential portion arises. Perplexity positions this as a first for agentic AI, and the rollout to users is now underway.