Apple positioning for agents
Apple appears to be positioning devices as the default home for agentic experiences by hosting third‑party models on‑device — social posts note Apple hosting options like Gemini locally — and Notion’s beta shows custom agents running in TestFlight builds. Those moves suggest more third‑party agent experiences could arrive tightly integrated with Apple hardware and native apps. ( )
Apple now runs Google's Gemini AI model directly on iPhones and iPads without cloud connections. This on-device hosting lets the AI process tasks locally for speed and privacy. (x.com) Users spotted the Gemini Nano model integrated into Apple's Shortcuts app in iOS 18.4 betas. Developers can now invoke it via app intents, treating it like Apple's own tools. (x.com) Notion's TestFlight beta for iOS also runs custom AI agents on-device using Apple's frameworks. These agents handle tasks like summarizing notes without sending data off the device. (x.com) AI agents are software programs that act autonomously, like a virtual assistant booking flights or editing documents without constant human input. They differ from chatbots by executing multi-step plans using tools such as calendars or browsers. (apple.com) Apple's Foundation Models framework, announced at WWDC 2024, lets third-party developers download and run AI models locally on Apple Silicon chips. Models like Gemini Nano—Google's smallest version with 1.8 billion parameters—fit within iPhone's Neural Engine limits. (machinelearning.apple.com) On-device AI avoids sending user data to remote servers, reducing latency to under 100 milliseconds for simple queries and enhancing privacy under laws like Europe's GDPR. Cloud AI, by contrast, risks data breaches and requires internet access. (apple.com) These betas build on iOS 18's Apple Intelligence features, launched in 2024, which already run models like OpenELM on-device. Hosting rivals' models expands the ecosystem without Apple building everything itself. (apple.com) Google confirmed Gemini Nano support for iOS in December 2024, calling it a step toward "multi-model AI experiences." Other firms like Anthropic may follow as Apple's framework gains traction. (blog.google) Microsoft tested similar on-device agents in its iOS Copilot app last year, but Apple's hardware edge—its Neural Engine hits 38 trillion operations per second—sets it apart from Android rivals. Qualcomm's chips lag at 45 TOPS across the ecosystem. (9to5mac.com) Rivals like OpenAI push cloud-based agents via apps like ChatGPT, which topped iOS download charts in 2025. Apple's moves could lock users into its hardware for seamless agent use, challenging that dominance. (sensortower.com) Watch for iOS 18.4's public release this spring, expected to formalize third-party model hosting. Developers predict a surge in native agent apps by WWDC 2026. (macrumors.com)
