Xcode 26.3
Apple (RC xip, downloads, Hacker News):
Xcode 26.3 introduces support for agentic coding, a new way in Xcode for developers to build apps, powered by coding agents from Anthropic and OpenAI. With agentic coding, Xcode can work autonomously toward a developer’s goals — from breaking down tasks to making decisions based on the project architecture, and using built-in tools to get things done.
In addition to Anthropic’s Claude Agent and OpenAI’s Codex integrations, Xcode 26.3 makes its capabilities available through the Model Context Protocol, an open standard that gives developers the flexibility to use any compatible agent or tool with Xcode.
For more information, see Setting up coding intelligence.
The agent sits in Xcode’s sidebar where developers can use it to plan new features, implement them, and review the results. As developers work, the agent generates a transcript of its actions, which lets developers follow along and interact with it. For example, code snippets will appear in the sidebar that can be clicked to take developers directly to the spot in the file where the agent made a change. Code updates can also be simultaneously previewed.
So Xcode just builds entire apps without you now
Xcode’s Codex support will happily trundle away for half an hour sticking its tendrils into every little corner of your project, touching and changing every file. It’s certainly going to be fun to build new projects with, but ain’t no way in hell I want to let that loose on any of my existing apps 😂
Xcode 26.3 getting official Claude and Codex integration without the usual guardrails is interesting enough, but having MCP in the mix is… unusually open for Apple.
[…]
But at least they seem to have done their homework where it regards the in-editor agent harness–not sure how deep they went into IDE primitives (file graph, docs search, project settings), though, and the apparent lack of skills and other creature comforts that all the VS Code–based tools have is a bit of a bummer.
I don’t know if this is super-duper interesting news, but I think it’s super-duper interesting that Apple saw the need to release this now, not at WWDC in June.
They couldn’t even wait for the final version to be shipping before sending out the press release.
I presume Apple announced these integrations now, and not at WWDC, to capture some of the frenzy surrounding tools like Cursor and Copilot.
As a heads-up, it’s around this time of year that Xcode traditionally goes new-OS-only, i.e. requires macOS 26.
They haven’t done that with Xcode 26.3 just yet, but you might find that this is the last point release to run on Sequoia.
AI models can access more of Xcode’s features to work toward a project goal, and Apple worked directly with Anthropic and OpenAI to configure their agents for use in Xcode. Agents can create new files, examine the structure of a project in Xcode, build a project directly and run tests, take image snapshots to double-check work, and access full Apple developer documentation that has been designed for AI agents.
If you’re an developer for Apple’s platforms and were wondering where you rank in their list of priorities consider that they were apparently capable of writing docs and adding meaningful Xcode integrations all this time but they decided to do it to help AI models instead of you.
This repository contains system prompts and documentation from Xcode 26.3, providing insights into Apple’s approach to AI-assisted coding and comprehensive guides for iOS 26 features and frameworks.
Combine is officially dead. Quote from Xcode 26.3 AgentSystemPromptAddition:
Avoid using the Combine framework and instead prefer to use Swift’s async and await versions of APIs instead.
It’s still not officially deprecated, but it’s obviously not preferred for new development.
Previously:
- Codex App
- Apple LLM Generating SwiftUI
- Xcode 26.2
- What Xcode 26’s AI Chat Integration Is Missing
- How to Use Google Gemini in Xcode 26 Beta
- Claude Code Experience
- Swift Assist, Part Deux
- Xcode + Claude
- Tim, Don’t Kill My Vibe
- Whither Swift Assist?
- How to Use Cursor for iOS Development
- Xcode 16 Announced
- LLM Coding Tools in Xcode
Update (2026-02-05): Dimitri Bouniol:
512 GB is a rough starting SSD size for development… All I did is install Xcode, my usual git repos, and a few other apps, and I’m already at >50% used before anything iCloud has even been touched…
Apple has honed in their
config.tomlto supercharge iOS development. Notes on Liquid Glass, call outs for Foundation Models — the list goes on. Though Apple blasts the doors off of their “big” stuff at W.W.D.C., you’d be crazy to think they aren’t paying attention. How we develop software is changing, and internally, it’s clear they are humming along with it. The fact that Xcode 26.3 exists, right now, is proof. They didn’t just cut a new branch once Codex’s macOS app shipped.[…]
So, how is the actual experience? Well, pretty nice! This is such a tiny thing, but in Terminal — removing a chunk of text sucks. I’m sure there is some keyboard shortcut I’m missing, or some other app I could use like iTerm or what have you, but not being able to use
Command+Aand then delete it hurts. In Xcode, that’s easily done because the input is not longer running through Terminal, it’s just an AppKit text entry control.[…]
This is a fantastic start for Xcode. If you’re later to the Claude Code or Codex scene, this is a wonderful place to start. There’s simply no going back once you learn how to use these tools. Ideas that you wanted to hack on become doable, those dusty side project folders come alive a bit more, and you get ideas out of your head much faster.
I handed it the classic SameGame codebase, gave it my coding style markdown file, and said “So this is an old ObjC app for iOS. I would like you to completely convert it 1:1 to modern Swift, with the coding style in mind. Leave no ObjC behind”
No other prompts needed; I needed to update a few legacy things in the xcode project settings (min OS version, Swift version, etc), and got this[…] All automatic, not a line of ObjC remains. Deprecated APIs were all modernized. 5,400 lines of ObjC became 2900 lines of Swift 5.
I really hate to say it, but at least on a test project, this Xcode 26.3 coding agent stuff is pretty good. I can see why some are going gaga over it.
My main concern (other than copyright, which for some reason companies seem to ignore) is that it makes it to easy to just trust the generated code. You start reading things line by line and after a few minutes its hard not to skim over the suggested bits that get added.
Update (2026-02-06): Christian Tietze (Mastodon):
For example, ever wondered when to use the new
InlineArray? See Swift-InlineArray-Span.md[…][…]
That is a very good summary that is painfully absent on the InlineArray API docs. As a Swift veteran, you usually look for a Swift Evolution proposal for the new tech then an try to find out there what this is all about.
[…]
The collection also has docs for 3D charts,
VisualIntelligence(I didn’t know that framework existed!), and UIKit Liquid Glass guides andAttributedStringtutorials.These documents are probably not written by a human, or team of humans, because of inconsistent tone and all. So I’d wager they were LLM-generated themselves. I do hope they were edited for misinformation at least!
Update (2026-02-09): Steve Troughton-Smith:
From what I’ve seen, Xcode/Codex is way better at creating great UIKit-or-AppKit-based apps than great SwiftUI apps, probably because there’s so much more source material to reference and train on, and hierarchies are more rigid and well-defined.
All of my attempts at having LLMs create SwiftUI apps ended up in frustration, but I’ve been playing around with a UIKit app for a couple of days now and getting very interesting results 👀
I think the syntax of SwiftUI itself is particularly confusing for LLMs. I’ve seen many models mess up things such as opening and closing braces when writing complex SwiftUI code. They also tend to bake too much business logic in view code instead of modularizing things, which doesn’t happen when doing UIKit.
Simultaneously, agentic AI tooling has trivialised the overhead of writing imperative layout boilerplate. Two key disadvantages of UIKit, slow development and verbose APIs, have been negated. The models are also trained on mountains of UIKit code and docs. Good luck one-shooting fresh SwiftUI APIs like TextEditor.
In 2025, the question is unavoidable:
Should you start migrating back to UIKit?
🛠️ Have you been using any of the new agentic tools in Xcode 26.3? What have you been using it for, and how has it fared?
I’m curious to hear everybody else’s stories, rather than just talk about mine all the time.
I’ve spent several hours today playing around with agents in Xcode 26.3. There are lots of things I like with the integration, but by far the most frustrating aspect is that the agents will sometimes just “forget” that they’re running inside Xcode and start trying to run terminal commands for things that they can do with the Xcode integration, even simple things such as reading content from files. The good thing is that telling them to use Xcode instead usually works.
Has anybody else seen the scrollbars being wildly off in Xcode/Tahoe?
I had a find operation, and even though there were only 3 occurrences, the scrollbar kept showing me moving down in the file even when it was actually wrapping around to the top.
Xcode 26.3 still can’t deal with you typing into a file while the file changes on disk without freezing (after displaying 2 alerts that block each other)
Update (2026-02-19): Two weeks later, the final version of Xcode 26.3 still isn’t available, and Apple has released an Xcode 26.4 beta.
So while this is definitely unusual, it’s not unprecedented. If history is any guide, I expect we’ll see the final release Xcode 26.3 within the next couple of days, and it will likely be identical to the Release Candidate.
Unsurprisingly, but still disappointingly, there’s been no update to the bundled version of Codex, which is stuck on gpt-5.2-codex
Update (2026-02-20): Xcode 26.3 RC 2 (xip, downloads):
Xcode 26.3 RC 2 includes Swift 6.2.3 and SDKs for iOS 26.2, iPadOS 26.2, tvOS 26.2, macOS 26.2, and visionOS 26.2.
Once again, the release notes don’t call out what’s new in this build, but diffing shows this major change:
Xcode 26.3 introduces a comprehensive permissions system that gives you fine-grained control over what external agentic coding tools can do in your development environment.
And lots of bug fixes, mostly related to AI.
I wonder if there's something weird about .3 releases… #Xcode 16.3, 15.3, and 14.3 all had two release candidates.
The RC 2 version of Xcode 26.3 does indeed distribute a new version of Codex that enables the gpt-5.3-codex model.
I would guess that’s why they waited before publishing 26.3 to the App Store, and it also answers the question as to how their model update strategy might play out.
Just noticed this new adaptive control in Xcode 26.4 😮
Update (2026-03-02): The final version has been released (xip, MacRumors).
Can someone at Apple vibe-code their way into this not being an app-modal window that prevents me from doing any work until it completes?
The more bonkers thing? It’s not modal for the copy. It’s modal for the run. Hit cancel and the copy continues.
it still happens in the background if you hit cancel
I am happy (?) to report that tabs in Xcode 26.3 are slightly less bad than they have been since 26.0. There are fewer situations that unintentionally create duplicate tabs of the same file (though still >0, which is still too many). Tab bar can now be scrolled instead of squishing every tab until they’re unrecognizable.
Still not as good overall as Xcode 16, but it’s progress. Good work, keep at it, folks.
Update (2026-03-04): Marco Arment:
Anyway if you ever get a wall of Xcode “Missing package product” errors, and no amount of cleaning/deleting/refreshing anything from the blog posts and AI agents actually fixes the problem…
It might be that one of those package’s Package.swift files got deleted.
13 Comments RSS · Twitter · Mastodon
The need for a truly curated app store just went up by a factor of a billion
https://mastodon.social/@stroughtonsmith/116019153868915403
Really interesting follow up from Steve Troughton-Smith
Interesting that Smith's conclusion is "RIP cross-platform frameworks."
Wouldn't you use the AI to generate cross platform code instead because ultimately that would result in less code for you to review? Unless we just aren't going to do that anymore.
I think it would be good to translate into native code rather than focus on cross platform code, to a degree. Isn’t that what we complain about already, lowest common denominator bloated electron apps because nobody can be bothered to do proper native Mac apps anymore?
Correct me if I’m wrong but if you’ve designed the core logic and assets and functionality is it so bad to let something convert it to native on multiple platforms than to do things like Catalyst? To use a simple example. Even Apple can’t seem to handle basic cross platform even as they define it within their own closed system.
Not to mention the Intel to AS transition. They gave the impression it’s as simple as checking some boxes and recompiling but I’ve been told it is not that simple. Maybe with this Apple can actually make it that simple. Especially since they’ve already announced Rosetta 2 is getting the axe.
> I think it would be good to translate into native code rather than focus on cross platform code, to a degree. Isn’t that what we complain about already, lowest common denominator bloated electron apps because nobody can be bothered to do proper native Mac apps anymore?
> Correct me if I’m wrong but if you’ve designed the core logic and assets and functionality is it so bad to let something convert it to native on multiple platforms than to do things like Catalyst?
I agree (or used to agree) that native apps are usually better (before SwiftUI).
> Even Apple can’t seem to handle basic cross platform even as they define it within their own closed system.
Exactly. But if you're using an LLM to write "native" for every target platform that theoretically increases the amount of code you have to review substantially. You have to review the Windows code, Linux, Android, etc. Now if you are using something like Flutter with your LLM you have a lot less code to review. I seriously doubt that the type of developer who will make an app mostly on vibes really cares about being "native." The claim that agentic coding is somehow the killer of cross platform frameworks is kind of interesting. I don't think so but I of course could end up being wrong.
@ObjC4Life - I was kind of thinking the same thing wrt to cross-platform development. Having to review gobs and gobs of LLM reports and code is still quite a bit of work. Having to do that on only one codebase is definitely a huge plus. I also agree that cross platform devs likely aren't going to care about native experiences all that much.
This might not be a popular opinion, but I actually find Flutter, Dart and their Cupertino widgets far more productive than Swift and SwiftUI. Dart is whatever enough that you don't really have to know much about it to use it. I haven't looked lately, but the last time I did, the LLMs were SUPER productive and accurate with Flutter and Dart versus what I saw with native Apple platforms.
Does it support converting unreadable Swift code (a.k.a. Swift code) to readable Swift code (or better like Obj-C code)?
Can it automatically create a feedback assistant ticket along with a simple project? Or make you save time by handling a discussion in the developer forum when it can not figure out how to use an API?
Apple employees, who claim on LinkedIn that no permissions are needed to be a developer, in order to brag about Xcode 26.3 should consider reading the agreement that is mandatory to accept to distribute apps on Apple’s platforms.
