A local AI agent still uses the same agent-loop pattern as a cloud agent — plan, call a tool, read the result, act again — but the orchestrator and, for on-device inference, the model itself run on hardware the user owns. Perplexity's Portable Computer is a shipping example: it runs entirely on a Windows or Linux PC with a qualifying NVIDIA RTX GPU (24GB+ VRAM for on-device inference), works with local files and connected apps like Outlook, Gmail, and GitHub without sending those tasks to the cloud, supports local MCP connections to a user's own tools, and can still fall back to frontier cloud models for harder reasoning — a hybrid local/cloud design rather than a fully offline one.