Hybrid agentic inference

The idea solves a three-way tension that cloud-only agents can't resolve cleanly: accuracy demands the most capable models, privacy demands that some data never leave the device, and cost demands you don't spend frontier compute on tasks a smaller model can handle. Hybrid agentic inference adds a routing layer that reads each incoming task or subtask, classifies it by sensitivity and complexity, and decides where it runs without asking the user to choose.

Perplexity named the pattern and shipped it in production on September 1, 2026, launching what it calls Hybrid Compute in the Perplexity Mac app. Under that system, a cloud agent can start a task and then hand off steps involving private files or sensitive data to a local model running on Apple silicon. An on-device PII classifier reads content before it leaves the machine, masking names and account numbers before any cloud call, then restoring them in the returned answer.

The pattern is gaining broader traction beyond Perplexity. Microsoft shipped Foundry Local in April 2026 for fully on-device inference on Windows, macOS, and Linux. Apple Intelligence routes sensitive processing to local M-series chips. For builders, hybrid agentic inference matters because it changes privacy and cost assumptions: your agent can handle sensitive documents without sending them to a cloud API, and it can reserve expensive frontier calls for steps that actually need them.