What WWDC 2026 changes if you build iOS apps with AI

The short version of WWDC 2026 for anyone building iOS apps: AI coding agents are now built into Xcode 27 (Claude, Codex, and now Gemini), Apple ships its own set of agent instructions that you can export into other tools, and SiriKit is officially on a deprecation clock, so App Intents is now the only way your app talks to Siri. That is the news. What the keynote cannot tell you is what building this way is actually like. I have shipped two iOS games with an AI agent in the loop the whole way, so this is the practical read, not a recap.

What Apple actually shipped

The agent story did not start this week. Apple quietly added agentic coding to Xcode 26.3 back in February, with Claude Code and Codex. What Xcode 27 does is finish the job:

• Gemini joins Claude and Codex as a built-in agent option, so all three major model families now plug straight into the IDE.
• Agents can check their own work. They can write and run tests, try ideas in isolation in Playgrounds, look at visual changes through previews, and drive a simulator or a real device through the new Device Hub. This matters more than it sounds: an agent that can verify what it did needs far less babysitting than one that just writes code and hopes.
• Code completion now runs fully on your Mac. Xcode 27's inline completion uses a model trained specifically for Swift and Apple's SDKs, running on the Neural Engine. Your source never leaves the machine for completions. The cloud agents are a separate, opt-in choice.
• Housekeeping that developers will feel: Xcode 27 is about 30 percent smaller, runs on Apple silicon only, and the old Simulator app is replaced by Device Hub.

On timing: the Xcode 27 developer beta went up on June 8, and the finished version ships alongside iOS 27 in the autumn, usually mid-September.

So Apple's pitch is a two-layer setup: private, local completion for the small stuff, and your choice of frontier agent for the big stuff.

Apple's agent skills are not locked to Xcode

Buried in the tooling news is the most immediately useful thing of the week. Xcode 27 ships with seven official agent skills: packaged instructions, written by Apple, that teach an agent Apple's own current guidance. There are skills for modern SwiftUI practice, migrating XCTest to Swift Testing, auditing your project's security settings, adopting the new SwiftUI APIs, and more.

The part most coverage missed: you do not need to move your workflow into Xcode to use them. Xcode 27 can export the whole set to a shared folder on your Mac, and tools like Claude Code, Codex, and Cursor pick them up from there on their own (SwiftLee has the full walkthrough). The one catch: this needs Xcode 27 itself, which is beta-only until September, so it is something to try on the beta now or bookmark for the autumn. Either way, it solves a real problem. Coding agents are good at Swift in general but shaky at the edges: brand-new APIs they have barely seen and freshly deprecated patterns they keep reaching for. Now the vendor's own guidance travels with the agent. The only caveat is that skills needing a running device or your project configuration still work best inside Xcode itself.

What agentic coding is actually like

Here is the part the keynote skips, from months of shipping this way.

The wins are real. An agent in the loop is the reason a one-person studio can ship two games in a few months. It writes the boilerplate, chases the build errors, and does in an afternoon what used to take me a week.

But agents are confident in both directions. The same setup that fixes a bug in minutes will also report problems that do not exist. When I ran a deep AI audit over one of my games, it came back with 13 performance findings. I checked every single one against the actual code. Exactly one was a real bug. Several of the proposed "fixes" would have quietly changed how the game looks. An agent that can run its own tests, like the Xcode 27 ones now can, catches some of this on its own. It does not catch the cases where the code is correct and the finding is imagined.

So the workflow that works is the same whether your agent lives in Xcode or a terminal: let it do the heavy lifting, verify what it claims, and write the corrections down somewhere it can find them again. That last part is the piece no vendor shipped this week. An agent with no memory of your project's hard-won lessons repeats the same wrong guesses forever. I keep mine in a searchable knowledge base the agent checks before it plans anything (the setup is in how to give your assistant a searchable memory, and the highest-value notes are the things it got wrong).

The deadline hiding under the AI news

The announcement that will actually bite app developers got the least airtime. The rebuilt Siri only sees apps through App Intents. SiriKit, the old framework, received a formal deprecation notice, with removal expected in roughly two to three years.

The trap is the timing. Old SiriKit code still compiles under iOS 27, just with warnings, so it feels safe. But apps that have not adopted App Intents are simply invisible to the new Siri from day one. The migration window is years; the visibility hit is this autumn.

If your app has any voice or Shortcuts surface, the move is: audit what you have on the old framework, pick the three to five actions users most want to trigger by voice, and ship those as App Intents first.

Quick hits worth knowing

• Core AI is a new framework for running open-source models on-device, with a curated set (Qwen, Mistral, and others) optimized for Apple silicon. Running a small local model next to a big cloud one is exactly how my own tooling works on the Mac, and Apple just made that pattern a first-class citizen on the phone.
• Foundation Models grew an abstraction layer: your app can swap between Apple's on-device model, Claude, or Gemini without changing the calling code.
• Swift concurrency keeps getting friendlier. The approachable defaults that took the pain out of Swift 6 adoption are now the baseline that Xcode 27 and the new SDKs assume, so strict concurrency hurts a lot less than it did a year ago.
• It was a big week for Claude generally. Anthropic released Claude Fable 5, its strongest model, the day after Apple demoed Claude agents in Xcode. The agent in your IDE is about to get noticeably smarter, and at 10 dollars in and 50 dollars out per million tokens, noticeably pricier when you reach for the top tier.

So is Xcode now the best place for agentic coding?

Apple said exactly that on stage, and the room full of developers on Hacker News had a good laugh. The honest answer: the integration is genuinely ahead in one way. No terminal-based agent can natively look at a SwiftUI preview, drive the simulator, and pull crash reports from the Organizer. For verify-your-own-work loops, the IDE has tools nobody else has.

Everything else still favors whatever workflow you already trust. The models are the same ones. The skills export means Apple's guidance is no longer locked to Xcode. And Apple has a recent history of demoing AI features a year or two before they ship properly, so I would treat the agent features as promising and check back when the beta settles.

It is not a switch-or-stay decision. Export the skills, keep your workflow, and let Xcode's agents earn their place on the tasks where device access matters.