Apple Built the Context Layer
Apple shipped a three-tier privacy stack that turns your messages, mail, and screen into agent infrastructure.
Ambient AI has a data problem. The agent needs to know your life - your messages, emails, calendar, what is on your screen - but sending that context to the cloud is a privacy disaster. Apple just shipped the architecture that solves this.
What Shipped
At WWDC on June 9, Apple unveiled Siri AI - a ground-up rebuild of its assistant as an autonomous agent with full personal context access.
The technical foundation is a three-tier privacy stack:
- On-device models handle sensitive queries. Messages, emails, photos, screen content - processed locally, never leaving the device.
- Private Cloud Compute runs moderate-complexity requests on Apple silicon servers with no data retention and verifiable code.
- A custom 1.2-trillion-parameter Gemini model handles complex reasoning, routed only when the lower tiers cannot resolve the query.
The routing is automatic. The user does not choose which tier processes a request. The system decides based on sensitivity and complexity.
The Architecture That Matters
Three changes make this more than a Siri upgrade.
Autonomous chaining. Siri AI executes multi-step tasks across apps without returning to the user between steps. Parse context, check calendar, search for options, book, send invites - one request, five apps, zero handoffs. This is agent orchestration at the OS level.
App Intents as the agent surface. Apple deprecated SiriKit and made App Intents mandatory. Every app that adopts the framework exposes structured actions and entities that Siri can discover and chain. This is MCP at the operating system layer - reaching two billion active devices instead of developer toolchains.
Foundation Models API. A native Swift API to run on-device models directly. Multimodal inputs, dynamic profiles that swap models mid-session, and pluggable providers including Claude and Gemini. You can build agent features that never touch a server.
What Changes for Builders
If you build ambient AI systems - and I do - Apple just shipped the context pipeline you were going to build yourself. Personal data indexing, semantic search across private data, multi-app orchestration, and a privacy architecture that routes sensitive queries to the right compute tier automatically.
The strategic question is the old one: do you build on Apple's context layer and accept platform dependency, or build your own pipeline and own the full stack? The bar for building your own just got significantly higher.
Where This Heads
The agent hardware shipped last week. The context layer shipped this week. What matters next is the shift from reactive to proactive - agents that act on what they sense without waiting to be asked. That transition is the distance between a better assistant and an actual ambient intelligence.