← Back to blog
WWDCSwiftSwiftUIAI

WWDC 2026 for devs (part 2): AI in apps, agents, and more SwiftUI

S
Slekens
Jun 14, 2026 · 3 min read · (updated Jun 14, 2026)

This is part two of the WWDC 2026 developer recap. In part one we covered the betas, the main SwiftUI changes, Swift 6.4, Xcode 27, and the AI APIs. Here we gather what came out over the rest of the sessions week, which mostly complements the AI-in-apps and agents angles.

Note: this is still preliminary information, from press and community coverage. API names and details may change during the beta.

Foundation Models: the LanguageModel protocol

What part one only hinted at with “Dynamic Profiles” now has a name: a public Swift protocol, LanguageModel, that cloud model providers implement to expose a common inference surface.

The practical upshot is the interesting part: you can switch AI providers by updating a Swift Package Manager dependency, with no changes to the session logic or the rest of the app. Google (Gemini, via the Firebase Apple SDK) and Anthropic (with a compatible Swift package) are already confirmed. The on-device model runs on the Neural Engine, and heavy queries escalate to Private Cloud Compute or the provider’s cloud. (TechTimes, Appcircle)

The idea, roughly:

// Your session code doesn't change; what changes is the dependency (the provider)
let session = LanguageModelSession()
let reply = try await session.respond(to: prompt)

LanguageModelSession stays the same; the concrete provider is decided by the package you link. (Conceptual names per coverage; the real API may differ during the beta.)

The framework also adds built-in tools: BarcodeReaderTool, OCRTool (via Vision), and a Spotlight-powered search for on-device RAG (retrieve local context and feed it to the model).

Xcode 27: dual-engine agents

We learned more about Xcode 27’s agentic system, and it’s dual-engine:

  • Local: a Neural Engine-tuned model gives real-time Swift suggestions specific to your project.
  • Cloud: routes heavier analysis to Claude (Anthropic), Gemini (Google), or OpenAI.

The agent can run tests, inspect visual changes in live previews, drive the Simulator through the Device Hub, and work with multi-turn streaming responses. It wires up 20+ tools via the Model Context Protocol (MCP). (TechTimes)

SwiftUI and SwiftData: the missing details

From Fatbobman’s technical write-up, a few details that didn’t make part one:

  • @State with lazy initialization.
  • Binding<Value?> (optional) support in alert and confirmationDialog.
  • A new DataDetection framework to recognize content (dates, addresses, etc.) inside Text.
  • The new document-based apps model boils down to: observable document objects + async snapshots + dedicated readers/writers.
  • SwiftData: @Attribute(.codable) as an escape hatch for opaque Codable types.

One honest caveat: custom transitions are still unavailable — for now, only the system ones (zoom and crossFade).

The mood of the year: pragmatic

The community’s overall read is that this WWDC was more about stability and efficiency than breakthrough features: limitations from past years were removed, gaps were filled (SwiftData, documents, toolbars), and there was a strong bet on agents and on-device AI. For everyday app building, that often matters more than a flashy headline.

Still developing. Once Apple publishes the official documentation and more sessions, we’ll verify the API names and expand. If you missed the basics, they’re in part 1.

👉 Keep going: part 3 is out — Xcode Agent Skills and the official SwiftUI APIs.

Sources: TechTimes — Foundation Models / providers · Appcircle — dev recap · Fatbobman’s Swift Weekly #139