Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

MacRumors

macrumors bot
Original poster


After several weeks of testing, Apple today released Xcode 26.3, an update that allows developers to use tools like Anthropic's Claude Agent and OpenAI's Codex for app creation directly in Xcode.

macOS-26-Xcode.jpg

With support for agentic coding, Xcode can complete more complex app development tasks autonomously. Apple worked with Anthropic and OpenAI to configure their agents for use in Xcode and to ensure that AI models can access a full range of Xcode features.

Agents are able to create new files, examine code structure of a project, build a project directly and run tests, take image snapshots to check their work, and access Apple's full, up-to-date developer documentation.

Along with OpenAI and Anthropic models, Xcode 26.3 is compatible with any agent or tool that uses the open standard Model Context Protocol.

Xcode 26.3 is available for download today from Apple's developer website.

Article Link: Apple Releases Xcode 26.3 With Support for AI Agents From Anthropic and OpenAI
 
Cool - this has been announced for a while. Glad it's finally out in the world and, while I'm not a software developer, I'm eager to give it a spin and a hobbyist coder. One step closer to full-on vibe coding with Xcode.
 
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%.
Yes - but for those of us who know a bit of code but aren't 'skilled' it opens up a new world of possibilities. I don't claim to be a coder - heck I only really know Python from a class I took a few years back and then have been tinkering ever since.

AI has been a game changer for me. Instead of scratching my head over syntax or commands, I get to express my idea, have AI write the code and then iterate in ways I just couldn't do before. I don't really care if it's great code or not as long as I get the results I need. I've built simple but time saving tools this way in ways in minutes instead of days or weeks.

An analogy for me would be - can a skilled barista make a better coffee than a super-automatic? Of course! But is a super-automatic an option for decent coffee if you don't have barista skills - hell yes!
 
This is the beginning of their "Anyone can build an app" marketing.

Guaranteed.

Not saying people are going to vibe code the next Legend of Zelda game, but the typical small business app requirements, a few APIs and a CloudKit integration? Easy peasy.

I'm a systems engineer. The entire role is Powershell and Bash. Claude codes those stone cold, without breaking a drop of sweat.
 
Yes - but for those of us who know a bit of code but aren't 'skilled' it opens up a new world of possibilities. I don't claim to be a coder - heck I only really know Python from a class I took a few years back and then have been tinkering ever since.

AI has been a game changer for me. Instead of scratching my head over syntax or commands, I get to express my idea, have AI write the code and then iterate in ways I just couldn't do before. I don't really care if it's great code or not as long as I get the results I need. I've built simple but time saving tools this way in ways in minutes instead of days or weeks.

An analogy for me would be - can a skilled barista make a better coffee than a super-automatic? Of course! But is a super-automatic an option for decent coffee if you don't have barista skills - hell yes!
Yep and I am a developer, just not an iOS or even Mac developer. But since getting Claude I’ve written my first Mac apps, signed up for the developer program and started writing iOS apps. These will never go on the App Store but I’m using them on my phone, solving use cases that I’ve had in my head for years.

Last weekend I built an app and backend that scrapes the commuter bus schedule and announces on my Sonos when a bus is 15 minutes away. That’s integrated with an MCP server that lets me tell Claude “I’m going to the office tomorrow” and it knows to set my Sonos alarm to the right time and start announcing buses the next morning. I wrote the iOS app to add a notification so it shows on my watch and keeps getting updates after I leave.

The Sonos part also tracks all the songs I play, including via SiriusXM and presents stats and a favorite function. Just last night I added MusicKit support so now that list of favorites syncs to my Apple Music as a playlist.

Nowadays if I see an app that does something I want but just slightly off, I just write a custom app exactly to my specs, and with no ads or subscriptions.
 
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%.

First, that study wasn't out of MIT. It was an independent group.

Second, you're oversimplifying what they found.

That study showed senior developers were slower on brownfield type tasks. When there is deep architecture, a particular project set of jargon/idioms, and large-scale dependencies, senior developers were slower with AI tools (that existed at the time of that study). However, if you are starting from scratch or doing more simple tasks, that same study found evidence that AI can increase productivity.

It's also important to understand the limitations of that study and what can be inferred from their data. From that paper: "The slowdown we observe does not imply that current AI tools do not often improve developer’s productivity—we find evidence that the high developer familiarity with repositories and the size and maturity of the repositories both contribute to the observed slowdown, and these factors do not apply in many software development settings. For example, our results are consistent with small greenfield projects or development in unfamiliar codebases seeing substantial speedup from AI assistance."

So experienced developers using mature repositories they are highly familiar with were faster than AI-assisted developers on particular tasks. For many other situations, there is solid evidence of speedup in other studies and in that one. Focusing just on the slowdown misses the bigger picture. Granted, the authors of the paper highlighted the slowdown more than the speedups, but they were pushing back against a broader set of data or perception that using AI tools always speeds up productivity. That is not the case.

But this study does not show developers are always slower with AI tools, just sometimes slower. Overall, the data suggest AI can slow down some developers in some areas and speed up in other areas.

I'll give an anecdote as someone who is not a developer. I've been writing Bash (yes, not a "real" language) code for a couple decades. I have some experience with Python and R. It's been many years since I've done anything with BASIC, C, C++, or Java.

What do I do? I'm a scientist who uses a lot of coding for my research. I now have Gemini and ChatGPT generate Python, R, and Bash code all the time for my work. It does exactly what I need in about 1/10th the time it used to take me to do things. That's not an exaggeration. If anything, it's an under-estimate.

Just a quick example. Getting data ready for my dissertation analyses involved me writing a 3000+ line script. That took some months to work through and verify everything. If that was the only thing I needed to work on, maybe I could have done it in 1-2 weeks. I can generate just as complex code that does similar things and much more (and with me verifying output!) within a day or two now.

It was a good learning experience for me at the time, but it also meant that instead of having time to work on other research that would lead to publications, I used all my time on my dissertation. Again, that was fine, but I appreciate being able to get through much more coding in much less time than I used to. I have more time to think about the science.

But I'm not a developer and my anecdote only applies to my situation.
 
Last edited:
If it’s like the RC then it’s tacked on. Not a huge benefit for the experienced and not enough hand holding for the novice. They are letting two entire companies be inside their app yet Apple is not lifting one finger to train the AI to get mac or iOS programming right. This isnt going to open the floodgates of complex apps, it will open the floodgates of slop buggy simple apps. Most of the real nice looking and behaving apple like stuff in apps is so convoluted to apply or you are severely limited by sandboxing its unusable in the same way an Apple designed app is. It’s going to take a few years and a few competitors to really outdo them in this space for them to catch up. They will offer AI services and automation on a subscription basis before they ever make it easy to build apps that don’t rely on pay me pay me pay me.
 
  • Like
Reactions: AF_APPLETALK
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%.
You are hallucinating details. The study wasn't done by/at MIT, it was done by METR. The number 22 isn't used anywhere in the paper as a percentage.

Also they have this disclaimer:
"We do not provide evidence that:
AI systems do not currently speed up many
or most software developers"
 
This is the beginning of their "Anyone can build an app" marketing.

Guaranteed.

Not saying people are going to vibe code the next Legend of Zelda game, but the typical small business app requirements, a few APIs and a CloudKit integration? Easy peasy.

I'm a systems engineer. The entire role is Powershell and Bash. Claude codes those stone cold, without breaking a drop of sweat.
AI coding is useless to inexperienced developers. However, it supercharges experienced developers, turning months of work into days. Literally.
 
I have yet to see a skilled developer write better code than AI. In particular, the SotA models released since Opus 4.5

That's a very ignorant and insulting thing to say about the developers whose decades worth of coding was used for training these models.

According to the benchmarks used in the AI communities and according to Anthroptic and OpenAI, the newest models have an 80% pass rate compared to a highly skilled developer.

The only thing advantage models have is speed, but that's because their responses are brute forced. Brute forcing an answer often results in splattershot coding that is too messy, not modular enough, and needs human feedback to track down all the bugs.
 
AI coding is useless to inexperienced developers. However, it supercharges experienced developers, turning months of work into days. Literally.

Maybe two weeks can be condensed into a day or two, not "months". But then there is a lot of time spent testing and debugging.

If what you said was true there would now be a a hundred Photoshop, After Effects and Cinema 4D clones vibe coded in the last year.

There's nothing remotely close to that. All we have seen is thousands of copycat menu bar apps and AI chat apps asking for subscription payments.
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.