Apple Releases Xcode 26.3 With Support for AI Agents From Anthropic and OpenAI

[AF_APPLETALK]

Xcode is a dog water ide so if I can just have an agent deal with it in the background then all the better

Xcode is a major liability for Apple, at this point. I hate using it. I used to love using it. They stopped developing it and advancing it like a decade ago, which is a major shame. It's in an absolutely pathetic state of existence compared to how it used to stack up against competitors, and how it's stacking up against competitors now.

This should be a top priority for Apple. If they lose developer, they're going to be back where they were wrt general adoption of their platforms in 2001.

I have yet to see AI write acceptable code and I've said it before and I'll say it again, AI slows down a skilled developer. MIT even did a study on it showing that developers get slowed down between ~19% - ~22%.

Not true.

Been using RC for awhile and the AI is terrible. Recommends code that does not compile. Makes up syntax. It is not worth the effort.

This is because the models aren't well enough trained on Apple's platforms, and because Apple's platforms are one GIGANTIC MESS and have been since they introduced Swift. Constant changes mean that the training the LLMs do have is often out of date. Even if you're an expert-level Apple developer, most don't know how to resolve these issues. Swift Data being a prime example.

The language is just not compatible with the system frameworks in the ways it needs to be, and it's too late an hour to reinvent the system frameworks from scratch (though that didn't stop them trying with SwiftUI (which is apparently losing political capital inside of Apple as it's failed to meet expectations)). It took decades of human capital and expertise and UI/UX research to deliver AppKit and UIKit. Then they just threw it all away with a rubberized C++.

Yes, Objective-C needed to go (or go through a MAJOR facelift). Swift was, however, the incorrect choice for a replacement, and now they are so ****ed beyond belief.

AI coding is useless to inexperienced developers. However, it supercharges experienced developers, turning months of work into days. Literally.

Can confirm.

Not true, I vibe coded an app in xcode. Totally functional, runs like a champ.

That's what you think presently. If you're not a software developer by trade, you have no idea how much work maintenance and adaptability is. Especially if you didn't reason about the system you built, to begin with.

Coming from Zed and not having used Xcode in a long, long time... I really hate it. So much.

We're at a midle stage in the AI grief process now, there's some bargaining happening in the zeitgeist now although many are still in the Anger phase. 4 more to go and it will take a year or more for a lot of people.

I can't stand Xcode. I have mainly coalesced on JetBrains IDEs now, they work damn good with Codex. Xcode feels like a mess. The AI functionality is bolted on and is too slow of a workflow compared to other AI tools. ****, it can't even manage to do a `git fetch` without crashing on my machine. I have to drop into the CLI and orchestrate everything that way when I working with Xcode repos. I repeat that it's absolutely pathetic.

 

[Green Valkyrie]

If AI was so good it wouldn't make up non-existent methods (constantly), it wouldn't get math wrong (constantly), it wouldn't get simple things like character counts wrong (constantly), it wouldn't churn out **** architecture.

Those of us who actually know what we're doing in software dev know AI code is ****. Pure ****.

18 months ago this was true and I gave up using it for that reason. The models have improved massively since then. Been a long time since I saw a hallucinated method. It did used to happen, and frequently, but not any more.

 

[irusensei]

I wish Apple stopped this this ridiculous sanctioned business partner approved corporate. They should allow you to set any compatible API endpoint.

 

[gank41]

I have yet to see AI write acceptable code and I've said it before and I'll say it again, AI slows down a skilled developer. MIT even did a study on it showing that developers get slowed down between ~19% - ~22%.

I've been able to get a relatively simple game made that works on my Mac, iPhone, and my TV, and I haven't written a single line of code yet. All Ai. It needs a lot of work, still, for it to be its own and what I want, but for how far I've been able to get with it already, I'm totally amazed.

When I get build errors, I'm using the Ai to tell it to fix them, and it asks how and explains what my options would do, and then when I tell it to do whatever that may be, it does. I've also been sticking with the free version of Claude. I can see, though, if I was further along with developing my game, that paying could be worth it. With a single prompt pass you can basically update your app to adopt all the new API's of whatever new OS is out. Saves SO much time..

 

[thefredelement]

I think a lot of these posts are fluff and lack important context. I use AI everyday, I've been writing code for over 10 years on Apple platforms.

In the last year the biggest change hasn't been model ability but apps (coding harnesses) around them, which essentially makes filling LLM context faster with better information.

LLMs still have context windows, still suffer service degradation and are limited to the data they've been exposed to - they're over biased to 'answer' what you asked it and that doesn't mean do it in a good way, it means make the user happy with the output. This varies based on the user's experience, I'm most effective with LLMs in one off, low context, mechanical speed up scenarios, compared to 'vibe coding' where I find the output is horrible, overly verbose, overly commented and will be hard to scale from.

 

[beermode]

That's what you think presently. If you're not a software developer by trade, you have no idea how much work maintenance and adaptability is. Especially if you didn't reason about the system you built, to begin with.

Lol. I've added several features, created a git, debugs fine, etc. It's extremely lightweight and it's absolutely heaven sent for our niche. I'm not saying developers are useless, but their gatekeeping won't save all of them. I wanted this app for years, and I was told I need two-three different developers and a months time just to get a working product. I did it a few days.
Good luck, gatekeeping on a public forum won't stop me from getting the apps and SaaS I want. Sorry if AI is a threat. It sucks, but don't lash out.

 

[CalMin]

Look at Nvidea vibe coding ))) reversed releases, bugs and terrible stability - it will be your future with vibe coding

I think we're talking about two different things. If you're talking about professional software development, then sure - all that stuff matters.

I'm talking about a hobbyist coding and small projects. Stuff i could do myself but that would take me (at my beginner skill level) days or weeks. I can have things running in minutes with AI. That's a game changer for me!

 

[ArrayDecay]

Attention arrogant, smug, and self-important neck-beards at Stack Overflow: you're finished.

 

[CATiNTHESTORE]

I think we're talking about two different things. If you're talking about professional software development, then sure - all that stuff matters.

I'm talking about a hobbyist coding and small projects. Stuff i could do myself but that would take me (at my beginner skill level) days or weeks. I can have things running in minutes with AI. That's a game changer for me!

It is all beginning from small steps and then you are in trap

 

[rezwits]

I have yet to see AI write acceptable code and I've said it before and I'll say it again, AI slows down a skilled developer. MIT even did a study on it showing that developers get slowed down between ~19% - ~22%.

I don't see how you can get slowed down when, you can walk away and double your time.
I feel like they did the study wrong.

Cause look, if I can spend 10 minutes writing up a prompt/setup/prep, and then paste it in and THEN, walk away, go brush my teeth, check on it, say that looks good but I need this. Go the thru the house empty the trash in the rooms and put the trash outside. Come back tell it again, "something else to add." Go back to kitchen, empty the dishwasher, top shelf, tell it "something else." Go back to kitchen, empty the dishwasher, bottom shelf, repeat, then silverware, then repeat. Then take care of the cat food. Go to the bathroom, etc.
Sit down and it's done! Play with the app for like 15-20 minutes and then fine tune fixing little bugs. I mean?

I doubled my time, it made my app while I spent 10 minutes writing prep, then 2-3 minutes 6 more times, I spent 30 minutes over a 90 minute period. I mean look, that's how I code now days... I feel like I just got 3 hours minimum done in 1.5 hours. etc etc It's doubling time at MINIMUM, if not 5x 10x for the work it's doing...

2¢

 

[Almost60]

seems to me if you know your industry then using industry words and industry specific framing will create results from AI using words that follow. Watch a lawyer and law student ask for same thing - good lawyer knows what to ask. But the thing about AI that people don't seem to utilize is ask what to ask. New tools work better after using. If you wrote the code then how do you know AI can't make it better. Maybe ask AI what you are not asking.

 

[BeatsByTim]

Xcode is a major liability for Apple, at this point. I hate using it. I used to love using it. They stopped developing it and advancing it like a decade ago, which is a major shame. It's in an absolutely pathetic state of existence compared to how it used to stack up against competitors, and how it's stacking up against competitors now.

This should be a top priority for Apple. If they lose developer, they're going to be back where they were wrt general adoption of their platforms in 2001.


Not true.


This is because the models aren't well enough trained on Apple's platforms, and because Apple's platforms are one GIGANTIC MESS and have been since they introduced Swift. Constant changes mean that the training the LLMs do have is often out of date. Even if you're an expert-level Apple developer, most don't know how to resolve these issues. Swift Data being a prime example.

The language is just not compatible with the system frameworks in the ways it needs to be, and it's too late an hour to reinvent the system frameworks from scratch (though that didn't stop them trying with SwiftUI (which is apparently losing political capital inside of Apple as it's failed to meet expectations)). It took decades of human capital and expertise and UI/UX research to deliver AppKit and UIKit. Then they just threw it all away with a rubberized C++.

Yes, Objective-C needed to go (or go through a MAJOR facelift). Swift was, however, the incorrect choice for a replacement, and now they are so ****ed beyond belief.


Can confirm.


That's what you think presently. If you're not a software developer by trade, you have no idea how much work maintenance and adaptability is. Especially if you didn't reason about the system you built, to begin with.


I can't stand Xcode. I have mainly coalesced on JetBrains IDEs now, they work damn good with Codex. Xcode feels like a mess. The AI functionality is bolted on and is too slow of a workflow compared to other AI tools. ****, it can't even manage to do a `git fetch` without crashing on my machine. I have to drop into the CLI and orchestrate everything that way when I working with Xcode repos. I repeat that it's absolutely pathetic.

In an ideal world they would have just adopted a competent language like c# instead of inventing new ********.

 

[onepotatoetoo]

We have "proper" agentic development workflows in mainstream products for less than a year, hold your horses. It's moving insanely fast.

Why aren't you trying to lecture me without knowing if I have used this "agents" for coding?

So instead of me holding horses, try holding yourself back from writing nonsense on the internet.

These agentic workflows are only useful in the hands of someone who knows the ins and outs of software development but even then they come with limitations.

For example, there is not one language model that can create a novel complex algorithm. Try it yourself and see with your own eyes how it will generate a known algorithm even when you ask it to make a new one. If it isn't in their training data they stumble. If Swift is updated to 6.3 tomorrow then it will take many months for the models to catch up even if they search online.

Also, if it was so brilliant then the developers at Blender would not have halted development of the iPadOS version of Blender. They would say "Hey ChatGPT port Blender to iPadOS. Here's the codebase. The hard work has already been done."

But what happened? AI couldn't do it. It would be a buggy mess that would take long time to debug and test because the developers would have to read and track down every commit.

That's the issue with coding with a language model. If it is not your code and you don't have the skills to read it, then it takes you much longer to debug and test.

 

[2128506]

"Hey ChatGPT port Blender to iPadOS. Here's the codebase. The hard work has already been done."

That is exactly what a clueless newbie would do - and then bitch about AI producing "bad code".

It's not AI issue, it's user issue.

 

[throAU]

I have yet to see AI write acceptable code and I've said it before and I'll say it again, AI slows down a skilled developer. MIT even did a study on it showing that developers get slowed down between ~19% - ~22%.

I have had AI write plenty of usable code, if you haven’t, maybe try some paid models and learn how to work within their limitations (i.e., get it to work on specific things that will fit in context).

LLMs are now at the point where it is often quicker and easier for me to vibe code my own utilities than go looking for them on the internet.

Code is now cheap enough to generate that its very much worth spending 15-20 minutes prompting an LLM to automate stuff and see how it goes before bothering to grind through some workloads.

 

[onepotatoetoo]

That is exactly what a clueless newbie would do - and then bitch about AI producing "bad code".

It's not AI issue, it's user issue.

OK you seem to have missed the conversation and maybe need a reminder.

More than a couple of people claimed, absurdly, that AI models produce better code than all human developers.

If that was the case then the Blender Foundation could indeed just say "Hey ChatGPT port Blender to iPadOS."

But it is not the case. Every successful use of AI coding for very complex applications require a lot of hand holding, inspection, reiteration and testing. Even more so than coding manually because if the machine is coding you don't know what is has done until you inspect it.

So coding time comes down, but inspection, debugging and testing times go up. For large companies that can get quite messy because now all the employees can blame an AI model when things go wrong. It's the ultimate 'the dog ate my homework' excuse. We have seen several examples lately of companies apologising for a model's output, not just in programming but in media too.

There is also an issue with what you said above.

"not AI issue, it's user issue"

Sure there can be users who don't know enough programming lingo in order to chat with the models properly, but you should never outright blame users over models that are known to do random things. When you do things like that you're acting like someone who worships a safetensors file and thinks humans should be annihilated from the planet.

 

[loopy123]

Attention arrogant, smug, and self-important neck-beards at Stack Overflow: you're finished.

So true, SO has seen their traffic drop like a stone lately. I’ve posted on SO, and it was so frustrating to be lectured at by those people.

 

[gank41]

I have had AI write plenty of usable code, if you haven’t, maybe try some paid models and learn how to work within their limitations (i.e., get it to work on specific things that will fit in context).

LLMs are now at the point where it is often quicker and easier for me to vibe code my own utilities than go looking for them on the internet.

Code is now cheap enough to generate that its very much worth spending 15-20 minutes prompting an LLM to automate stuff and see how it goes before bothering to grind through some workloads.

I've been able to get pretty far along with creating a new game, only using Ai Prompts with the free version. When people complain about Ai here, it's not from actually using it in Xcode. It's probably some other reason they hate Ai.. TBH, there's little I'm needing to correct, and if there's any build issues, the Ai can fix that, too. So far this game works on my Mac, iPad, and Apple TV.. Just having different issues installing it on my Phone but that's a user error on my end, not anything to do with Ai..

 

[rezwits]

With agency Xcode can build debug write error debug build build write build debug write error debug build ad nauseam! And if you worked with Xcode in the past week with GPT 5.3-codex and Claude. You probably have noticed we are strictly talking about tokens and time...

The statement "It's only a matter of time" has never been truer. I have seen it. I mean it.

When you first experience this in Xcode you have a giggle that's sicker than when you first saw Chat GPT 4.1 make a website from scratch.

You create the project and write a laundry list like:

I want a project that does this <2-3 sentences>. (and then):
Please start with this and proceed doing the following steps:
– (1st)
– (2nd) ...
– (12th)

If you meticulously layout the order in an Apple Framework and Logical way, it has NO PROBLEMs.

You walk away and in 30-45 minutes (after an initial "approve all" of CMDs)
When you get back (it's done the debug build loops all done) and hit the "run" button in Xcode and it's just there the most AMAZING feeling. It's DONE. All you have to do is like the old website days of GPT 4.1, you have do the "Interior Design" i.e. the Graphical User Interface (Pleasantries). And if you want you can use AI to design graphics!

Some will say "Yeah and `that's the part that sucks about AI` hahah, because you know it can't flesh out the GUI 100% (these guys always go to the "future" for their complaints, yeah but the anti-gravity car there's too much traffic, "I am waiting for teleportation" hahah), but with SwiftUI and all cross-platforming, I mean what do you want?

But the fact that it just keeps going and going... oh amazing...