I bogged me that it says that he spend loads of time upfront planning but not with cursor. You can do the same with cursor, and on my experience, they are about the same.
When they reach certain level of complexity of a medium sized project they start messing up.
I agree. It’s just not a fair comparison. But then the article is written by Kino with the author “sprinkling in” sentences, so I guess I can’t really expect much.
I tried out Kiro last week on quite a gnarly date time parsing problem. I had started with a two hundred-ish word prompt and a few bits of code examples for context to describe the problem. Similar to OP it forced me to stop and think more clearly about the problem I was trying to solve and in the end left my jaw on the floor as I saw it work through the task list.
I think only early bit of feedback I had was in that my tasks were also writing a lot of tests, and if the feedback loop to getting test results was neater this would be insanely powerful. Something like a sandboxed terminal, I am less keen on a YOLO mode and had to keep authorising the terminal to run.
> This long (hopefully substantial) blog post was written collaboratively, with Kiro handling the bulk of the content generation while I focused on steering the direction.
So Kiro wrote whatever Kiro "decided", better said, guessed, what to write about and did most of the "content generation" - a weird but fitting term to use by a machine in writing a fake human blog. And the human kind of "directed it", but we dont really know for sure, because language is our primary interface and an author should be able to express their thoughts without using a machine?
I'd happier if the author shared their actual experience in writing the software with this tool.
Very good write up. This is indeed a qualitative change in the work of software engineering.
What they’re describing can also be done with Claude Code, and it’s way too broad in scope to get any benefit at all from approving code before it’s written. These tools are the way for now.
Cursor is in my opinion not geared for this level of hands-off.
Been using it for a while, now all other tools feel lacking. The multistep process with manual refinement options after every step just works so much better for keeping LLMs in line. The only times I’ve used anything but kiro since, has been when the Sonnet 4.0 was once again overloaded (tried it with 3.7 for a bit, but it went way off the rails of kiro’s tasks structure every time).
Kiro’s main advantage is Amazon is paying for my LLM usage instead of me.
For the most part it’s unlimited right now. Vs Code’s Copilot Agent mode is basically the same thing , tell it to write a list of tasks , but I have to pay for it.
I’m much happier with both of these options, both are much cheaper than Claude Code.
IMO the real race is to get LLM cost down. Someone smarter than me is going to figure out how to run a top LLM model for next to nothing.
This person will be a billionaire. Nvidia and AMD are probably already working on it. I want Deepseek running on a 100$ computer that uses a nominal amount of power.
Spec based development is a game changer for people with non-coding background to work on side-projects. I’m using this to fork a design flow I use for analog/RFIC design, and I can finally mend together open-source CAD tools to attune to the design flow, that I’d otherwise use in segregation.
I am just gonna say it. This is not something Kiro came up with. People were already using this workflow. Perhaps they should’ve added more features instead of spending time making promo videos of themselves. I fail to see any add value here especially considering it’s Amazon. Sonnet 4 is effectively unlimited for many MAX users so giving that away to work out their list of bugs is a non-starter.
I wanted a tiny helper tool to display my global keyboard shortcuts for me on macOS. I gave Kiro a short spec and some TypeScript describing the schema of the input data.
It wrote around 5000 LOC including tests and they... worked. It didn't look as nice as I would have liked, but I wasn't able to break it. However, 5000 lines was way too much code for such a simple task, the solution was over-engineered along every possible axis. I was able to (manually) get it down to ~800LOC without losing any important functionality.
at this point i think we ought to start having a tag in the submission title for when the submission is (primarily) llm-generated, like we do for video/pdf/nsfw
I'm long on no-AI becoming a label of quality, a badge of honor, as people are more exposed to the slop LLMs forces upon them.
I want prose, music, photography and movies made by humans, for humans.
I want to use software built by developers. The last part in particular because I've seen my share of bug riddled code spewed by the LLM du jour, I'm convinced our industry is vibecoding itself into the abyss.
> What I found interesting is how it forced me to think differently about the development process itself. Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
This is what I already do with Claude Code. Case in point, I spent 2.5 hours yesterday planning a new feature - first working with an agent to build out the plan, then 4 cycles of having that agent spit out a prompt for another agent to critique the plan and integrate the feedback.
In the end, once I got a clean bill of health on the plan from the “crusty-senior-architect” agent, I had Claude build it - took 12 minutes.
Two passes of the senior-architect and crusty-senior-architect debating how good the code quality was / fixing a few minor issues and the exercise was complete. The new feature worked flawlessly. It took a shade over 3 hours to implement what would have taken me 2 days by myself.
I have been doing this workflow a while, but Claude Code released Agents yesterday (/agents) and I highly recommend them. You can define an agent on the basis of another agent, so crusty-architect is a clone of my senior-architect but it’s never happy unless code was super simple, maintainable, and uses well established patterns. The debates between the two remind me of sitting in conf rooms hashing an issue out with a good team.
Super interesting that no one here has mentioned replit.
I found the experience of using Kiro and replit really similar with one important difference: replit mostly worked.
Kiro tore off and wrote tonnes of code and tests. It asked me to peck at approval requests (one thing I liked was the regexp type trust this tool to do this requests) it spent half a day creating an app to do what I asked... and it was incomprehensible bullshit. I couldn't do anything with the project and I have not touched it since.
Replit was a bit more interactive, but pretty autonomous, and it got me to 90% of the solution and then stalled out - wouldn't correct some of the problems I identified to it. About 2 hrs with Cursor sorted that out though.
I did use Cursor to do it "AI assisted" and that took about the same amount of time. The advantage is that I really do know whats going on in the code base, but the Replit + Cursor solution is actually better in the sense that it looks better, and works better because the agent did some bits a bit more nicely than I did with Cursor - so I got those ideas for free.
Anyway :
Hand coding = a walk through the wilderness
Cursor = motor cross scrambler bike up the mountain
Replit = a helicopter ride to somewhere higher up the mountain selected at random that you didn't know about but now you have to get to the peak by yourself buddy, good luck
Kiro = you are blindfolded in a container of some sort and it's moving.
> What I found interesting is how it forced me to think differently about the development process itself.
> Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
I don't want to sound rude, but isn't this is something that happens to you with experience after some years?
How can you even be a senior developer without "spending more time articulating what I actually wanted to build and high level software architectural choices"?
I use https://sketch.dev, which seems to be approximately the same kind of thing. I find it is very organized, it does not need nearly as much babysitting as this guy is talking about. Perhaps a difference is the "todo read" and "todo write" tools, which it uses to break the main goal into achievable tasks, and then judge itself on if it has finished correctly or not. Also, Sketch is easy-going with corrections: if you tell it that it went the wrong direction, it quickly adjusts and follows your direction.
Anyway, two big thumbs up for agentic dev workflows from me, but only so far with Sketch.
21 comments
[ 2.6 ms ] story [ 46.8 ms ] thread> With Kiro, I spend more time upfront articulating what I want to build, but then I can step back and let it execute
This sounds like exactly the kind of exercise one does to /init a project with Claude, define tasks/spec etc.
I think only early bit of feedback I had was in that my tasks were also writing a lot of tests, and if the feedback loop to getting test results was neater this would be insanely powerful. Something like a sandboxed terminal, I am less keen on a YOLO mode and had to keep authorising the terminal to run.
So Kiro wrote whatever Kiro "decided", better said, guessed, what to write about and did most of the "content generation" - a weird but fitting term to use by a machine in writing a fake human blog. And the human kind of "directed it", but we dont really know for sure, because language is our primary interface and an author should be able to express their thoughts without using a machine?
I'd happier if the author shared their actual experience in writing the software with this tool.
This blog post does approximate my sentiments pretty well, although its writing style diverges from my usual style.
What they’re describing can also be done with Claude Code, and it’s way too broad in scope to get any benefit at all from approving code before it’s written. These tools are the way for now.
Cursor is in my opinion not geared for this level of hands-off.
Can't really get value out reading this if you don't compare it to the leading coding agent
For the most part it’s unlimited right now. Vs Code’s Copilot Agent mode is basically the same thing , tell it to write a list of tasks , but I have to pay for it.
I’m much happier with both of these options, both are much cheaper than Claude Code.
IMO the real race is to get LLM cost down. Someone smarter than me is going to figure out how to run a top LLM model for next to nothing.
This person will be a billionaire. Nvidia and AMD are probably already working on it. I want Deepseek running on a 100$ computer that uses a nominal amount of power.
It wrote around 5000 LOC including tests and they... worked. It didn't look as nice as I would have liked, but I wasn't able to break it. However, 5000 lines was way too much code for such a simple task, the solution was over-engineered along every possible axis. I was able to (manually) get it down to ~800LOC without losing any important functionality.
It sounds like a very PM type approach to coding.
Does that mean it fits PM types more than IC dev types?
> What I found interesting is how it forced me to think differently about the development process itself. Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
This is what I already do with Claude Code. Case in point, I spent 2.5 hours yesterday planning a new feature - first working with an agent to build out the plan, then 4 cycles of having that agent spit out a prompt for another agent to critique the plan and integrate the feedback.
In the end, once I got a clean bill of health on the plan from the “crusty-senior-architect” agent, I had Claude build it - took 12 minutes.
Two passes of the senior-architect and crusty-senior-architect debating how good the code quality was / fixing a few minor issues and the exercise was complete. The new feature worked flawlessly. It took a shade over 3 hours to implement what would have taken me 2 days by myself.
I have been doing this workflow a while, but Claude Code released Agents yesterday (/agents) and I highly recommend them. You can define an agent on the basis of another agent, so crusty-architect is a clone of my senior-architect but it’s never happy unless code was super simple, maintainable, and uses well established patterns. The debates between the two remind me of sitting in conf rooms hashing an issue out with a good team.
I found the experience of using Kiro and replit really similar with one important difference: replit mostly worked.
Kiro tore off and wrote tonnes of code and tests. It asked me to peck at approval requests (one thing I liked was the regexp type trust this tool to do this requests) it spent half a day creating an app to do what I asked... and it was incomprehensible bullshit. I couldn't do anything with the project and I have not touched it since.
Replit was a bit more interactive, but pretty autonomous, and it got me to 90% of the solution and then stalled out - wouldn't correct some of the problems I identified to it. About 2 hrs with Cursor sorted that out though.
I did use Cursor to do it "AI assisted" and that took about the same amount of time. The advantage is that I really do know whats going on in the code base, but the Replit + Cursor solution is actually better in the sense that it looks better, and works better because the agent did some bits a bit more nicely than I did with Cursor - so I got those ideas for free.
Anyway :
Hand coding = a walk through the wilderness
Cursor = motor cross scrambler bike up the mountain
Replit = a helicopter ride to somewhere higher up the mountain selected at random that you didn't know about but now you have to get to the peak by yourself buddy, good luck
Kiro = you are blindfolded in a container of some sort and it's moving.
> Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
I don't want to sound rude, but isn't this is something that happens to you with experience after some years?
How can you even be a senior developer without "spending more time articulating what I actually wanted to build and high level software architectural choices"?
Anyway, two big thumbs up for agentic dev workflows from me, but only so far with Sketch.