Hi, I am one of those skilled people who uses the models and all I can say is I’m spending less time coding and more time screaming at the model for lying and more time testing for bugs. The overall production time is about the same, but with the stress of talking to a computer ghost.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Xcode 26.3 Lets AI Agents From Anthropic and OpenAI Build Apps Autonomously
- Thread starter MacRumors
- Start date
- Sort by reaction score
Hi, I am one of those skilled people who uses the models and all I can say is I’m spending less time coding and more time screaming at the model for lying and more time testing for bugs. The overall production time is about the same, but with the stress of talking to a computer ghost.
Exactly this. There is a lot of boilerplate code that AI can do well and much faster.
I am amazed by the tech, it’s impressive for sure. I do know how to use it. I, like a lot of other devs get it forced down our throat at every corner.
I am acutely aware of how it makes you dumber. Not just as an anecdote (of which I have many), as a scientifically studied phenomenon. Yes, even for using it for things you can already do blindfolded.
Also, personally, I don’t like it.
There I said it. It writes **** code because it’s trained on a mountain of **** code. The time it takes to make well structured guides for it, with code examples and good prompts, I’d rather just write the frigging code myself.
By doing so you activate parts of your brain which make you a better developer. By not doing so you actively damage those parts of your brain.
And to top it off - you won’t remember why you wrote the class that way in 6 months, or what reason you chose one reason over the other.
It’s pants, I’m sorry.
So you are now a manager, but of agents instead of people. 🙂
It depends a lot on what you are trying to do with the model. It can't cover 100% of use cases, and there are times where it will be stumped and just try random "fixes" (hallucination) stage. This is where experience will win out.
For boilerplate or "investigate this codebase", it is a big time-saver.
Maybe it will help those lazy iOS devs to finally transition their apps to Liquid Glass. I'm done with the inconsistencies, especially with the keyboard
You will, document it in markdown files so the AI can read it. Then in 6 months time, you ask "why is the class structured like this?" and it will tell you. It's akin to keeping documentation yourself - which many devs don't do!
You can write them yourself if you fancy yourself a better writer than a developer. Or just have it update the doc, it beats having no documentation at all.
This could be a bellweather of sorts. The Xcode number is 26.3. macOS 26.3 is scheduled to be released soon. Same with Xcode, coming next week.
Perhaps this is a sign we could get the M5 Pro MacBook Pro soon?
Perhaps this is a sign we could get the M5 Pro MacBook Pro soon?
App development is just undergoing it's AutoCAD moment. Nobody programs CNC code anymore as its all generated by machine paths from software. CNC code is pretty basic in comparison to something like Swift or Kotlin but it's exciting to see these tools gain a GUI and become more accessible as a result.
The job of machinist however didn't go away; it just went from maker to more of a supervising role.
The job of machinist however didn't go away; it just went from maker to more of a supervising role.
You probably think that made you sound like an avantgarde tech genius living in the future but it doesn’t. You’re still going to be drowning in AI generate vibe coded garbage and the general trend for quality software will go downhill even faster.
You’re basically watching the Gangesfication of software. A once holy river now full of stinking trash.
The Xcode LLM interface leaves a lot to be desired so far, hopefully that gets updated soon. Even still with 26.3, you have an LLM interface for each project where using Claude Code or codex in the desktop app lets you work across projects from one window. Though you do have to go the apps.
LLMs are great at what the video shows, stubbing a feature, cross referencing with docs, sometimes fixing its own build errors but still, regardless of harness or orchestration, struggles with complexity, which is to be expected with so much built in constraint around context. The weather feature addition was ~400 lines of code that claude hallucinated on during the first pass.
I find my best LLM time savings still comes from aim small miss small flows, mostly in single files, as a speed up. It's sometimes good with debugging but most of the time it's better to use it to add logging and then use it as a rubber duck compared to 'fix this bug'.
I've experimented with the 'cupertino' MCP (apple docs, swift evolution forums), skills, plugins, getting Claude Code to iterate with codex, etc. its exciting but no matter the setup the weakness will aways be the model.
Software engineering requires a high degree of context, most projects are specialized in various ways. I think until we can get LLMs that are trained on projects themselves (and continually updated) it's just going to be what it is.
I saw a YouTube short that had a quote like: "Some days I worry that AI is going to take my job, today is not one of those days" - which is a sentiment I agree with, though every time a new model or tool comes out I get a little twinge of, is this it? Are we finally at the jumping off point? I don't think we're there quite yet. I don't think it's going to take our jobs, if you're decent at what you do, you'll just use different tools than you did before. I don't know about you guys but my backlogs lately seem to be getting bigger.
LLMs are great at what the video shows, stubbing a feature, cross referencing with docs, sometimes fixing its own build errors but still, regardless of harness or orchestration, struggles with complexity, which is to be expected with so much built in constraint around context. The weather feature addition was ~400 lines of code that claude hallucinated on during the first pass.
I find my best LLM time savings still comes from aim small miss small flows, mostly in single files, as a speed up. It's sometimes good with debugging but most of the time it's better to use it to add logging and then use it as a rubber duck compared to 'fix this bug'.
I've experimented with the 'cupertino' MCP (apple docs, swift evolution forums), skills, plugins, getting Claude Code to iterate with codex, etc. its exciting but no matter the setup the weakness will aways be the model.
Software engineering requires a high degree of context, most projects are specialized in various ways. I think until we can get LLMs that are trained on projects themselves (and continually updated) it's just going to be what it is.
I saw a YouTube short that had a quote like: "Some days I worry that AI is going to take my job, today is not one of those days" - which is a sentiment I agree with, though every time a new model or tool comes out I get a little twinge of, is this it? Are we finally at the jumping off point? I don't think we're there quite yet. I don't think it's going to take our jobs, if you're decent at what you do, you'll just use different tools than you did before. I don't know about you guys but my backlogs lately seem to be getting bigger.
AI is useful but it also churns out a lot of slop.
If you vibe code, and don’t understand the implementation it’s going to get more difficult to debug issues that AI cannot fix. So more difficult to add features and maintain.
Also less secure - AI won’t harden the app for you - that is - blissfully unaware of possible vulnerabilities if you don’t know what to protect against.
If you vibe code, and don’t understand the implementation it’s going to get more difficult to debug issues that AI cannot fix. So more difficult to add features and maintain.
Also less secure - AI won’t harden the app for you - that is - blissfully unaware of possible vulnerabilities if you don’t know what to protect against.
You're not wrong, but there is a more nuanced approach here that objectively/empirically verifiable by anyone.
There are apps hundreds of millions of people use every day that have 100% of code at this point written by "agents" that are performant, quality, and enjoy high user sat levels.
There are also, for every one of those, an inordinate amount of terrible, useless apps getting worse b/c people are using the tools incorrectly.
Siri, is also, a complete unmitigated disaster. But none of that means that, wielded properly, AI-produced code cannot create an amazing product. It does so every day, but it's just like anything else: are we wielding the tools we have correctly? Obviously Apple has not with Siri, and that's incredibly embarrassing for them, but obviously many companies have figured out how 100% agentic code can work beautifully.
Except people vote with their wallet, there's always been a lot of bad code but most of it isn't successful. You're either B2B and 'good enough' (which can be very broad as far as quality), or you're B2C and unless you're truly unique, people don't pay for crap, at least not for very long.
No, I'm a developer, not a vibe coder. It has nothing to do with pretentiousness. Yes, I am well aware there is subpar quality software made by "AI bros" looking to make a quick buck off non-technical people. That is all over social media.
This is a tool, you can choose to ignore it and then when corporate mandates it, you throw the tantrum with them. Or you accept it, use it to further your career and land a better job. This also includes cybersecurity, as there will be no shortage of those jobs given how bad most vibe coded apps are.
So, now instead of just doing the thing, which I enjoy doing, I need to:
Make huge amounts of guidelines and code examples, architecture explanations - coupled with careful prompting to get this model to *maybe* spit out acceptable code. Then I need to review said code. Fix the performance, security, and other bugs of that code. And now I need to document it so *it* can remember why it did that code?
No thanks. I’ll just do it myself.
As for documentation - in my book that’s a smell.
Your code, including a consistent architecture and design patterns, coupled with descriptive method naming- is all the documentation you need.
Completely disagree. Code is not self-documenting, especially not to stakeholders who are not developers.
Can they please integrate this into iOS so that I can ask for new features and bug fixes and have them be directly implemented?
Those AI data centers eventually need to be payed off, you know.
I'm a software engineer that uses AI every day. But it's a lot like having a sewing robot. You can use the robot make clothes but that doesn't mean you can be Chanel. Also, the robot makes a lot of mistakes. It's great if you're smart, not great of you're not.
Aha! So you're the guy that's been hard coding IP addresses in all the apps I support that are having such monumental problems migrating to new infrastructure and also to the cloud! A few REM statement here and there, buddy, would have made it a lot easier for the devs to go back and fix. There's also this thing called DNS that comes in really handy too! 😉