Cursor introduces its coding model alongside multi-agent interface

The vibe-coding IDE put an emphasis on speed with Composer.

See full article...

 

[nzod]

IDE?

Isn't it rather an Integrated Vibing Environment?

 

[PilotlessDrone]

Cursor understands syntax so much better than I do, it can refactor, document functions, and introduce subtle bugs ten times faster than I ever could. So ... I like it but we have our moments.

 

[Missing Minute]

that means using multiple models at once for the same task and comparing their results, then picking the best one.

People are already doing this, the amount of tokens used is immense and it's likely going to keep increasing.

 

[wrosecrans]

As a person with the credibility of decades of experience, and at least verifiably a very old Ars account, I'd like to take a moment out of my day so that some of the younger folks can get some wisdom from a greybeard who has been through many industry hype cycles and has a capacity to put this current AI product release cycle into a sensible context. My analysis is thus,

Poop, poppie poop. Poopity poop. Poopity dirty poop, pooped from a butt. Butt that did how poop goes.

And thus, I have addressed the topic with precisely the substance it deserves, and greatly more substance than comes from any of the hucksters trying to pump their investments before the bubble pop. (Which will sadly not pop soon enough.) May all of this be buried, and the charlatans be damned.

 

[nzod]

I have been programming for more than half a century.

If you do not know who Grace Hopper is then look her up. Grace Hopper would be very excited by AI. Her vision for COBOL was for a language that non-programmers could understand. COBOL is often criticized for being wordy but it was designed to be. Now with AI people can write specifications for applications and those specifications can be understood by non-programmers. AI can generate applications from specifications that non-programmers can understand.

In the past management has avoided allowing programmers time to develop good specifications and documentation. They usually say they want good specifications and documentation but they just do not want to pay for it. In the future management will understand that good specifications will save money.

In the near future programmers will be separated into those that use AI and those that do not. We shall see which of those are more successful.

I raise you Ada Lovelace, who I personally know would hate the idea of not knowing intimately what, how and why you're creating. "What if the AI infrastructure is down?", I can hear her asking in her distinctive voice.

 

[OtherSystemGuy]

In the future management will understand that good specifications will save money because software can now be generated from specifications.

While having specifications is important, if those specs aren't written in a formal language/logic you cannot prove they're right or complete nor if there is any conflict. Until you can get specs without ambiguity, no automated system is going to generate complete, known good software.

 

[chaos215bar2]

I raise you Ada Lovelace, who I personally know would hate the idea of not knowing intimately what, how and why you're creating. "What if the AI infrastructure is down?", I can hear her asking in her distinctive voice.

(I assume people are aware she lived in the 1800s, and I'm just not quite getting the joke / sarcasm.)

 

[fixthesite]

I have been programming for more than half a century.

If you do not know who Grace Hopper is then look her up. Grace Hopper would be very excited by AI. Her vision for COBOL was for a language that non-programmers could understand. COBOL is often criticized for being wordy but it was designed to be. Now with AI people can write specifications for applications and those specifications can be understood by non-programmers. AI can generate applications from specifications that non-programmers can understand.

In the past management has avoided allowing programmers time to develop good specifications and documentation. They usually say they want good specifications and documentation but they just do not want to pay for it. In the future management will understand that good specifications will save money because software can now be generated from specifications.

In the near future programmers will be separated into those that use AI and those that do not. We shall see which of those are more successful.

The goal is not enabling programmers. The goal is eliminating them. Capital is always looking to reduce labor costs. Less specialized labor means you have a larger labor pool you can pay less. The false promise that AI sells is a repeat of the Industrial Revolution, where a specialized labor class was forced out of existence. Management exists to serve capital. They will not be more understanding. If anything they will demand more.

 

[dropadrop]

Out of my team it’s the most creative and talented developers that are doing pretty amazing stuff with GenAI.

I don’t really know about the juniors though. When I studied at university it was pretty strict to not use any libraries or even built in ”advanced classes” for a long time. Afterwards it’s clear why. Like our skilled developers, you maybe need to understand how something is done to effectively command something else to do it for you and then validate it actually works. But if you are already at a 1 in 1000 level then the real power is in your mind, the tool just saves you time.

 

[Missing Minute]

AI1 Prompt: What shoudl I do here?
AI2 Prompt: What shoudl I do here?
AI3 Prompt: What shoudl I do here?
AI4 Prompt: What shoudl I do here?

AI5 Prompt: Which of the above 4 outputs is the best course of action?

AI1 Prompt: Take the selected output and execute.
AI2 Prompt: Take the selected output and execute.
AI3 Prompt: Take the selected output and execute.
AI4 Prompt: Take the selected output and execute.

AI5 Prompt: Which of the above 4 outputs is the best execution?

AI1 Prompt: What are the flaws with the above output?
AI2 Prompt: What are the flaws with the above output?
AI3 Prompt: What are the flaws with the above output?
AI4 Prompt: What are the flaws with the above output?

And so on...

Surely this will be the future of programming!

 

[BobbyBobberson]

The goal is not enabling programmers. The goal is eliminating them. Capital is always looking to reduce labor costs. Less specialized labor means you have a larger labor pool you can pay less. The false promise that AI sells is a repeat of the Industrial Revolution, where a specialized labor class was forced out of existence. Management exists to serve capital. They will not be more understanding. If anything they will demand more.

This will exist in any economic system.

 

[Missing Minute]

edit: nevermind, this troll doesn't deserve feed.

 

[Fatesrider]

I have been programming for more than half a century.

If you do not know who Grace Hopper is then look her up. Grace Hopper would be very excited by AI. Her vision for COBOL was for a language that non-programmers could understand. COBOL is often criticized for being wordy but it was designed to be. Now with AI people can write specifications for applications and those specifications can be understood by non-programmers. AI can generate applications from specifications that non-programmers can understand.

In the past management has avoided allowing programmers time to develop good specifications and documentation. They usually say they want good specifications and documentation but they just do not want to pay for it. In the future management will understand that good specifications will save money because software can now be generated from specifications.

In the near future programmers will be separated into those that use AI and those that do not. We shall see which of those are more successful.

Are you really comparing two entirely different corporate focus eras and thinking that AI can help that? AI costs money. So do programmers. AI is wrong a lot. So are programmers. It's about money. It's always about money these days.

So, what others have said is largely true. Since AI and programmers are often wrong, as long as one is cheaper to use than the other, they'll go with that. It's not about fulfilling the dreams of long dead developers. It's not about what's "best". It's not about efficiency. It's about making money. AI costs money. Less than a programmer, too. And they'll go with whatever is cheaper. As soon as they decide the time is right to pull that trigger.

Will they keep BOTH? Probably not. That's paying (almost) two to do one job, and at an additional cost.

So which is cheaper: Pay for a developer to get the specification mostly right, and have the capacity to fix it if they notice it's screwed up, or have the AI do it all by itself and risk having egregious errors pushed to customers because the AI can't recognize mistakes?

This isn't a multiple choice. They will get rid of whichever is more expensive. How do I know? READ YOUR POST. What I wrote is the exact path the corporate world hast taken in the last 50 years. That's your proof for you of what the future holds. It's NOT about PRODUCT. It's about PROFIT. For some reason, you've not learned that you now considered a liability to a corporation, and you must prove your worth every day, or you get kicked to the streets. That's what they do today. THAT is the corporate mentality of our world.

Oh, yeah, and remember, they use AI's because AI's are cheaper, so that's really the only choice they'll make. Developers are generally toast. They're just beginning to feel the heat. Assuming AI doesn't blow up (which is not a safe assumption), they'll be cinders in short order. The few who remain will be cage cleaners for AI slop, fixing what they can't.

This is the core principle of "enshitification". And THAT, at least you should have noticed by now. It's probably kind of shocking to realize how much one's vocation is in the path of the corporate steamroller instead of someone else's in the organization. Developers are now a liability. AI isn't any better, and probably worse, but they're cheaper. And if corporations can continue to push AI-slop to customers and make profits from it, they will.

It's not a complicated math to do.

 

[wernsey]

A friend told me yesterday that their company had to let one of their programmers go recently because his excuse for not getting his work done on time was "I ran out of tokens"

I'm sure we're going to see more of that in the future

 

[Missing Minute]

A friend told me yesterday that their company had to let one of their programmers go recently because his excuse for not getting his work done on time was "I ran out of tokens"

I'm sure we're going to see more of that in the future

That would be funny if it weren't so sad.

 

[blookoolaid]

I've been a bit reluctant to adopt AI as a coding tool. My boss has been pushing me to use it more and over the last few months I've gotten used to using ChatGPT as a replacement for searching Stack/MDN/Google. Then I started using it as an aid for doing peer reviews. Then for creating project proposals. And even as a rubber duck for brainstorming.

After watching him "vibe" code with Cursor I was again skeptical. But today I decided to try using Cursor instead of writing the code myself. I had to make moderate UI changes to a mature app to match spec that was determined in recent meeting. I gave it an image to start with and told it where to put the buttons. I gave it some guidance on styles but mostly it just picked it up from the codebase.

To my surprise, it worked really, really well. And being able to switch between agent (doing stuff) plan mode (explaining what it would do in response to a prompt) made it feel okay to trust the agent. And it was good at backing out changes that didn't work out. Then I realized it could do commit messages for me....

I still don't know how I feel about it, but I can't deny that I did my regular job today via prompts instead of writing code and it was objectively a better, more productive experience.

 

[Sprigganmaster]

I feel A.I. would be ok if only experienced programmers would be allowed to use it. I know what I want, need and how to achieve a result.

Unfortunately, I see a lot of young devs using it blindly and they don’t realize that they are not building their knowledge by searching, questioning and making some mistakes.

It’s also sad because without building their knowledge, they are not becoming relevant and essential. They will be easy to replace as soon as Ai becomes better

 

[vaibhavsdalvi]

I feel A.I. would be ok if only experienced programmers would be allowed to use it. I know what I want, need and how to achieve a result.

Unfortunately, I see a lot of young devs using it blindly and they don’t realize that they are not building their knowledge by searching, questioning and making some mistakes.

It’s also sad because without building their knowledge, they are not becoming relevant and essential. They will be easy to replace as soon as Ai becomes better

IT guy with 15 years in embedded product development. Started using AI six months back, and see this same thing. It has increased my productivity by 30%. But my juniors see 30% fall when they try to use AI. Most of the time AI leads them to a rabbits nest of unsolvable problems or reaches an error loop where it fixes one error to produce a new one, fixing that brings back another different error, fixing that via AI we end up with original error. If AI keeps getting better and better, then with enough computing power it will one day replace organic intelligence.

 

[dollyllama]

[Not getting work done on time because he ran out of tokens] would be funny if it weren't so sad.

It's funny because it's an utterly lame excuse, because if he wasn't going to buy more tokens himself, he should've asked for more; but if he was really burning through tokens and not getting work done, then he's likely a lousy developer.

 

[chillbert]

While having specifications is important, if those specs aren't written in a formal language/logic you cannot prove they're right or complete nor if there is any conflict. Until you can get specs without ambiguity, no automated system is going to generate complete, known good software.

What percentage of successful software products need to be this "complete and known"? I suspect a low single-digit percentage. For most (but not all) software development needs, less-than-complete specs + testing + iterative improvement are plenty good enough and allow for value to be delivered without requiring exhaustive specification. And let's not forget that exhaustive, formally precise specs can still be wrong because the problem was not properly understood at time of writing.

 

[chillbert]

As a person with the credibility of decades of experience, and at least verifiably a very old Ars account, I'd like to take a moment out of my day so that some of the younger folks can get some wisdom from a greybeard who has been through many industry hype cycles and has a capacity to put this current AI product release cycle into a sensible context. My analysis is thus,

Poop, poppie poop. Poopity poop. Poopity dirty poop, pooped from a butt. Butt that did how poop goes.

And thus, I have addressed the topic with precisely the substance it deserves, and greatly more substance than comes from any of the hucksters trying to pump their investments before the bubble pop. (Which will sadly not pop soon enough.) May all of this be buried, and the charlatans be damned.

I have the privilege of working with several very experienced developers (20-30 years professional experience in enterprise software development) who are now getting significant value out of AI for coding and architecture development. Even my own limited experience as an amateur coder is that code explanation alone is a remarkable tool. But having seen the best coders in my company using it daily to gets things done under their expert direction, I am convinced that there is real value there. This can be true at the same time as the fact that there are plenty of charlatans in the business side of AI. I can only suggest you investigate further.

 

[NewCrow]

So far, a non-representative sample of developers I’ve spoken with has told me they feel that Composer is not ineffective, but rather too expensive, given a perceived capability gap with the big models.

So how much does it cost?

 

[cleek]

The goal is not enabling programmers. The goal is eliminating them. Capital is always looking to reduce labor costs. Less specialized labor means you have a larger labor pool you can pay less. The false promise that AI sells is a repeat of the Industrial Revolution, where a specialized labor class was forced out of existence. Management exists to serve capital. They will not be more understanding. If anything they will demand more.

"The underlying purpose of AI is to allow wealth to access skill while removing from the skilled the ability to access wealth."


View: https://www.reddit.com/r/Snorkblot/comments/1o4oq9g/its_the_natural_order/

 

[mdrejhon]

My preferred use of AI in any VSCode-derived IDEs is to baby step me through new programming techniques, or as a code reviewer. That's my new best-practices, as I have to keep up with industry (for better or for worse) even without vibing.

In any sidebar with AI chat, teach me on the fly in-IDE things like "How can I do this in WebGL"? And "ELI5 what is a vertex array"? To things like "Give me 3 ideas how to optimize this draw queue, including speed and better garbage collection management?" (As it looks at my code and give me relevant lessons customized to my learning style for my situation). Yes, these are the kind of things you can use an external AI for (whether you use Mistral or Claude or ChatGPT website) but an in-IDE AI does automatic awareness of multi-file code context.

Then I take over the coding myself, not vibing. I am a long time developer and I need to use my grey matter. So more teach / review / plan tasks than vibe tasks.

Also as a code commit reviewer. And weirdly, GPT 5 Codex Thinking seems to catch more coding errors in my git diffs than Claude 4.0 and 4.5 (even Thinking version). I inserted temporary intentional errors in my code changes and Claude missed them while Codex caught them AND found other of my mistakes that I overlooked.

Claude is faster and more "interesting code training teacher" to talk to... but has become too much of a "sychophant rubber stamp code reviewer" unexpectedly -- it thinks lots of my commits are good while Codex criticizes me. Claude is a more fun code teacher but Codex is a much better "code changes reviewer" on git diffs. It even caught a retroactive typo in my variable. I do have to ask it to fully repeat review my code changes, then it does the whole sandboxed "git diff" thing again to see my pre-commit changes.

I definitely WANT to be criticized for any of my code errors, so it's interesting GPT5-Codex (not normal GPT5) is the best criticizer. Take a longer time (5 minutes instead of 1 minute) figuring out if a potential code error is really a code error or false positive (or needs better documentation / variable rename to prevent confusing the AI or other humans). I can do something else while you're taking merry time doing superior-reviewing-job of my commit, but I want to know about my bugs rather than being sychophantly rubber stamped.

I want to do my own debugging. I really hate how AI IDEs no longer prioritized certain things. Cursor, for example, has a bug for months that don't let me do middle-of-line breakpoints like stock VSCode can. (Like the middle conditional of a if statement). It tries to do so, but doesn't let me.

(See the grey circles -- I want to turn those red, to breakpoint only on this.getWidth() as an example. But Cursor does not let me like VSCode can!)

Since I don't like my regular Classic Intellisense Tab modified -- I've turned off the single-Tab AI code autocompletes (although I have reassigned it to double-tab (Tab+Tab). So it prioritizes on Microsoft Classic Intellisense first, but if I glance good AI skeletoning, I may do the double Tab+Tab to accept/review (like redundant tedium typing / skeletoning / etc). Basically code I was going to type anyway, simply as a typing speed accelerator. This reduces the temptation to vibe code because I don't want to accidentally accept AI code changes when I was simply tabbing through my code, etc.

Don't let me atrophy my coding skills, please, in these IDE prioritizations to vibecode. If I must use AI help, let me use AI sparingly my way, let me organically code my way.

 

[NobodyImportant]

As a person with the credibility of decades of experience, and at least verifiably a very old Ars account, I'd like to take a moment out of my day so that some of the younger folks can get some wisdom from a greybeard who has been through many industry hype cycles and has a capacity to put this current AI product release cycle into a sensible context. My analysis is thus,

Poop, poppie poop. Poopity poop. Poopity dirty poop, pooped from a butt. Butt that did how poop goes.

And thus, I have addressed the topic with precisely the substance it deserves, and greatly more substance than comes from any of the hucksters trying to pump their investments before the bubble pop. (Which will sadly not pop soon enough.) May all of this be buried, and the charlatans be damned.

I've been a bit reluctant to adopt AI as a coding tool. My boss has been pushing me to use it more and over the last few months I've gotten used to using ChatGPT as a replacement for searching Stack/MDN/Google. Then I started using it as an aid for doing peer reviews. Then for creating project proposals. And even as a rubber duck for brainstorming.

After watching him "vibe" code with Cursor I was again skeptical. But today I decided to try using Cursor instead of writing the code myself. I had to make moderate UI changes to a mature app to match spec that was determined in recent meeting. I gave it an image to start with and told it where to put the buttons. I gave it some guidance on styles but mostly it just picked it up from the codebase.

To my surprise, it worked really, really well. And being able to switch between agent (doing stuff) plan mode (explaining what it would do in response to a prompt) made it feel okay to trust the agent. And it was good at backing out changes that didn't work out. Then I realized it could do commit messages for me....

I still don't know how I feel about it, but I can't deny that I did my regular job today via prompts instead of writing code and it was objectively a better, more productive experience.

Yes. And if this improves 2x year over year a few times, typing code will be an anachronism. I bet there is a 4x improvement available just from tooling integration and language design - even if the models totally stagnant in ability.

 

[TimeToTilt]

I have the privilege of working with several very experienced developers (20-30 years professional experience in enterprise software development) who are now getting significant value out of AI for coding and architecture development. Even my own limited experience as an amateur coder is that code explanation alone is a remarkable tool. But having seen the best coders in my company using it daily to gets things done under their expert direction, I am convinced that there is real value there. This can be true at the same time as the fact that there are plenty of charlatans in the business side of AI. I can only suggest you investigate further.

Here's my issue with this... Are they getting value or do they just think they are. Because so far all studies done by non industry money return no benefits or active negatives. Or in the case of this study where experienced devs thought they were going much faster with AI while actually being much less productive in reality.

https://meincmagazine.com/ai/2025/07/...source-software-developers-19-percent-slower/

 

[WXW]

I've been a bit reluctant to adopt AI as a coding tool. My boss has been pushing me to use it more and over the last few months I've gotten used to using ChatGPT as a replacement for searching Stack/MDN/Google. Then I started using it as an aid for doing peer reviews. Then for creating project proposals. And even as a rubber duck for brainstorming.

After watching him "vibe" code with Cursor I was again skeptical. But today I decided to try using Cursor instead of writing the code myself. I had to make moderate UI changes to a mature app to match spec that was determined in recent meeting. I gave it an image to start with and told it where to put the buttons. I gave it some guidance on styles but mostly it just picked it up from the codebase.

To my surprise, it worked really, really well. And being able to switch between agent (doing stuff) plan mode (explaining what it would do in response to a prompt) made it feel okay to trust the agent. And it was good at backing out changes that didn't work out. Then I realized it could do commit messages for me....

I still don't know how I feel about it, but I can't deny that I did my regular job today via prompts instead of writing code and it was objectively a better, more productive experience.

IT guy with 15 years in embedded product development. Started using AI six months back, and see this same thing. It has increased my productivity by 30%. But my juniors see 30% fall when they try to use AI. Most of the time AI leads them to a rabbits nest of unsolvable problems or reaches an error loop where it fixes one error to produce a new one, fixing that brings back another different error, fixing that via AI we end up with original error. If AI keeps getting better and better, then with enough computing power it will one day replace organic intelligence.

Meanwhile, my tests of AIs doing peer reviews have always been terrible (the AIs see a lot of issues which don't exist, and miss actual ones).

And a few weeks ago I scorned the project lead for using AI for the backend without review or even testing (and things obviously failed), and he said from now on he would ask for my review for any changes, so I recently did a review of some AI-made unit tests (which suppossedly AIs are good at...?), and they were shit. The cherry on the cake is that at first the AI didn't "want" to acknowledge the issues...

So I calculate that AI would cost me a 42.1673% of productivity, using the same estimation method done by other posters (i.e. whatever it feels like), and I would be miserable having to spend my time reviewing code.

 

[tenaku2]

While having specifications is important, if those specs aren't written in a formal language/logic you cannot prove they're right or complete nor if there is any conflict. Until you can get specs without ambiguity, no automated system is going to generate complete, known good software.

We already have specs without amiguity. It's called "code".

 

[BabelHuber]

I've been a bit reluctant to adopt AI as a coding tool. My boss has been pushing me to use it more and over the last few months I've gotten used to using ChatGPT as a replacement for searching Stack/MDN/Google. Then I started using it as an aid for doing peer reviews. Then for creating project proposals. And even as a rubber duck for brainstorming.

After watching him "vibe" code with Cursor I was again skeptical. But today I decided to try using Cursor instead of writing the code myself. I had to make moderate UI changes to a mature app to match spec that was determined in recent meeting. I gave it an image to start with and told it where to put the buttons. I gave it some guidance on styles but mostly it just picked it up from the codebase.

To my surprise, it worked really, really well. And being able to switch between agent (doing stuff) plan mode (explaining what it would do in response to a prompt) made it feel okay to trust the agent. And it was good at backing out changes that didn't work out. Then I realized it could do commit messages for me....

I still don't know how I feel about it, but I can't deny that I did my regular job today via prompts instead of writing code and it was objectively a better, more productive experience.

Do I understand this correctly? You made minor UI changes without changing the underlying logic of the program?

Sorry to say, but this sounds rather trivial. Okay, I can use Vibe for minor UI adjustments. But I can also spend a few minutes doing this by hand.

But as soon as a developer really has to get creative to implement complex algorithms, it's another story.

I had projects where I literally spent weeks to understand the requirements of the client, because these guys were unable to voice their requirements coherently. They told me something different every day.

And during implementation, still requirements were added as the customer "forgot" some things. Little details like half his supply chain. I kid you not.

Have fun using LLMs for such scenarios. I'm looking forward reading about those failed projects and their cost to the affected business.

 

[butcherg]

Hmmm, another layer of abstraction...

I'm all good with using a tool for leverage, but I still want to understand the lower levels because that's where computing things are done. I recall having to vector some of my students to the EE digital design class because I couldn't get an instructor for the CS equivalent. Most groused, but quite a few also came back later and expressed some gratitude for having to chump through the hardware stuff, relating how it helped them zero in on defect root causes quicker because they understood more about what was going on in the registers and such. Yes, a run-on sentence, so read it carefully,...

 

[icwhatudidthere]

I had projects where I literally spent weeks to understand the requirements of the client, because these guys were unable to voice their requirements coherently. They told me something different every day.

And during implementation, still requirements were added as the customer "forgot" some things. Little details like half his supply chain. I kid you not.

But these seem like things where "vibe coding" could help out. Vibe coding could really be called rapid prototyping with a natural language interface. Take a look at what Figma Make and Jetbrains Matter are doing. You could sit with a client and have rapid prototyping sessions and they can start seeing application flows and maybe start identifying missing parts before implementation starts.

 

[EspressoMachine]

That would be funny if it weren't so sad.

Depends also on how the work was setup. If it was set at such a rate that only a machine can do then this guy just shot himself in the foot by introducing the usage of AI.

 

[CUclimber]

So many people just miss the benefits of AI tools. Sure it's crap if a non-programmer just floods it with vague specs and expects a fully mature App out of it.

But if you know what you want to build and tell it very concrete specs on what to do and how to do it, it can save massive amounts of time. For example: I had a feature request on a site a few weeks ago that needed to add a custom watermark to each page of a PDF whenever someone clicked a link to view it. It needed to take an existing PDF file, add the watermark (with a bunch of user-based context) to every page, upload that file to an existing S3 bucket, then have the frontend request a signed URL to download/view the file; all of this necessitated a change from a synchronous front-end request to an async one. I also needed to implement some smart caching so that only one watermarked file per day per user would be generated, and perform cleanup/archiving of old generated files.

Any engineer reading this knows that what I described isn't terribly complicated, but it would take a few hours or a day to build and test. Claude did it in under 5 minutes, including the time it took to type out the instructions. It got it all done on the first try and saved me an entire morning of writing API endpoints, updating the front-end slightly, writing unit tests, etc. Massive, massive time-saver because I knew what I wanted and just had Claude build it for me.

All of the infrastructure was already in place, including the API backend, a front-end framework, AWS S3 and IAM roles etc. That was all put together manually at the start of the project so I know it's done right, and having all of your frameworks and tooling in place makes it far less likely that your AI agent will just start making up entire systems that shouldn't exist.

Far too many programmers are way too cynical about these tools. If you're smart about it they are an unbelievable advantage if you're trying to ship features quickly. Obviously check its work since it will occasionally recommend doing something stupid, but in my experience lately it gets it right far more often than it gets it wrong.

 

[BabelHuber]

But these seem like things where "vibe coding" could help out. Vibe coding could really be called rapid prototyping with a natural language interface. Take a look at what Figma Make and Jetbrains Matter are doing. You could sit with a client and have rapid prototyping sessions and they can start seeing application flows and maybe start identifying missing parts before implementation starts.

In trivial cases maybe. But when developing complex algorithms without a template for the algorithm - which is called "creativity" - you cannot. You can only do again what someone else already did.

 

[San Diego Dude]

Running multi-agent on the same repo is just asking for trouble. Multi-agent concept is neat in theory, but if you have > 1 agent working in the same environment, have fun cleaning up those conflicts. These things still aren't good enough to just 'turn loose' without having to go back over it and make sure it's not breaking things (at least if you're doing it right and not putting total faith in vibe coding), so the multi-agent concept right now is more science project than anything right now.

 

[CeeTee]

While having specifications is important, if those specs aren't written in a formal language/logic you cannot prove they're right or complete nor if there is any conflict. Until you can get specs without ambiguity, no automated system is going to generate complete, known good software.

I would suggest that specifications themselves are useless, without understanding the problem needing to be solved, and the impact of solutions on the business/entity implementing them. Developers are, first and foremost, interpreters of solutions to problems, and this first step is where so-called citizen programmers typically struggle.