cd ~/
Agentic AIPlatform ArchitectureAgent ArchitectureOn-Device AIAmbient AI

Apple Shipped the Context Pipeline I Was Going to Build

Apple shipped a three-tier privacy stack that turns your messages, mail, and screen into agent infrastructure.

Evyatar Bluzer
3 min read

Ambient AI has a data problem. The agent needs to know your life - messages, email, calendar, whatever is on your screen - but sending that context to the cloud is a privacy disaster, and every team building in this space ends up hand-rolling some partial answer to it. Apple just shipped a full one.

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 the sensitive queries: messages, emails, photos, screen content, all processed locally, never leaving the device. Private Cloud Compute takes moderate-complexity requests on Apple silicon servers with no data retention and verifiable code. Above that sits a custom 1.2-trillion-parameter Gemini model for complex reasoning, routed in only when the lower tiers cannot resolve the query. The user never picks a tier. The system routes each request on its own, based on sensitivity and complexity.

Siri AI's three compute tiersA request box at the top feeds a highlighted router box, which fans out to three boxes below: on-device models on the left, Private Cloud Compute in the middle, and custom Gemini on the right, each branch labeled with the condition that routes to it. A requestwith personal contextRoutersensitivity x complexity(the user never picks a tier)On-device modelsmessages, email, photos, screen contentnever leave the devicePrivate Cloud ComputeApple silicon servers, no data retention,verifiable codeCustom Gemini, 1.2T parameterscomplex reasoning sensitive moderate complexity only when the lowertiers cannot resolve it
Siri AI's three compute tiers and the router that picks one per request, as announced at WWDC on June 9: sensitive data stays on the device, and only what the lower tiers cannot resolve reaches the 1.2-trillion-parameter Gemini model.

What actually changed under the hood?

A smarter Siri would be a footnote. Three of the changes go well past that.

Autonomous chaining. Siri AI executes multi-step tasks across apps without returning to the user between steps: parse context, check the calendar, search for options, book, send invites. One request touches five apps and the user never sees a handoff. Apple is doing agent orchestration at the OS level.

App Intents as the agent surface. Apple deprecated SiriKit and made App Intents mandatory, so every app that adopts the framework exposes structured actions and entities that Siri can discover and chain. If you have built MCP servers, this will look familiar - same idea, moved down to the operating system layer, reaching two billion active devices instead of developer toolchains. In April I called the protocol debate settled in MCP's favor. At the OS layer Apple just declined to join it, and two billion devices give that refusal weight.

Foundation Models API. A native Swift API to run on-device models directly, with 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 platform question. Build on Apple's context layer and accept the 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 and the context layer shipped this week. What I am watching for now is the move 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.

Comments