This month marks the one-year anniversary of my journey into “vibe coding.” Looking back, 2025 will undoubtedly be remembered as the year the paradigm shifted with respect to software development, at least for me.
I have spent the last twelve months burning billions of tokens, shipping a dizzying amount of software, and obsessively using tools like Replit, Cursor, GPT Codex, and Claude Code, usually with a healthy dose of curiosity as my main driver.
The big question on everyone’s mind is: Is it here to stay? Was it a fundamental shift for software and the world?
In a sentence, I feel agentic (vibe) coding is the most significant change in software since the compiler.
I struggle to find a historical analog for what this shift has felt like. It’s a bit like the printing press, in that it automates something arduous. But unlike the printing press, the automobile, or the internet, the pace at which this technology is proliferating and improving, specifically agentic coding, is unlike anything I’ve ever seen. It’d be like November of 2024 the Wright Brothers took their first flight and In December we’re already on the moon.
For me, the feeling is less “industrial revolution” and more “Ironman.” In the course of a year, I was given powers I never thought I’d have in my lifetime.
Here is what I built, what I used, and what I think we are all getting wrong about the future of software.
The Stack: My “Vibe Coding” Arsenal
I’ve tried them all. Here is how the tools shook out after a year of heavy use.
1. The Daily Driver: Cursor
Verdict: The “Full Manual” Professional Camera.
Cursor has become my main IDE by a large margin. It is essentially a fork of VS Code that MIT alums turned into the fastest company to reach $1B in ARR.
If Replit is like shooting with a high-end camera on “Aperture Priority” (great results, automated settings), Cursor is shooting on “Manual.” You have total control. You can build anything, OS native apps, iOS apps, Linux tools, machine learning pipelines…you name it.
I love it because it allows me to run agents in parallel, pick any model I want (including local ones), and manage complex contexts. It was powerful enough to take me from a blank screen to a fully fundraising-ready iOS app.
2. The First Love: Replit
Verdict: The “Prosumer” Powerhouse.
Replit was my gateway drug. It sits in a perfect middle ground between professional tools and consumer toys. It is incredibly capable, especially for web apps.
What I love:
- Speed: I can spin up a landing page or a CRUD app in seconds.
- Integrations: Their integration with Neon (Postgres) is flawless. Spinning up prod/dev databases is a cinch.
- Safety: It keeps you on the rails. It’s hard to break things permanently.
Editor’s Note: Two days after publishing this, Replit announced they now support native iOS and Android development. I have not played with this, but its a significant improvement for the platform.
What I don’t:
- Cost: It can get expensive quickly compared to Cursor. I’m sure VC is subsidizing a lot of my Cursor usage, but there is no two ways about it, I spend a lot more with Replit for the same project than in Cursor.
- Constraints: You hit a wall when you want to build native mobile or OS apps. (this is no longer true)
- Capability: I still can’t build native MacOS software, something I’m doing more and more as I build custom ML/AI programs for unique use cases.
I still use Replit for prototyping and simple web tools, but for the heavy lifting, I’ve graduated to Cursor.
The Rest of the Field
- Claude Code: Amazing technology, but I don’t love living in the Command Line Interface (CLI). I prefer the UI features of an IDE. In a way, I’m still using Claude the most (just via Cursor). Yes, I know, its “better” via Claude Code, and cheaper, but…see above, I’m not a CLI guy.
- GPT Codex: A solid runner-up, but I mostly access it as a model within Cursor rather than a standalone tool. Also, again, not a CLI guy…
- Lovable: A cool product, but it occupies the same mental space as Replit for me, so it hasn’t stuck in my workflow.
The Output: What I Actually Shipped
The volume of output this year was insane. Most of what I’ve shipped you can find here. Sure, some of it was “slop,” but a lot of it is real software used by real people.
- Rali: My crown jewel this year. A complex iOS app that lets people spin up grassroots races. I iterated on this for over a thousand hours, moving from a concept to a functional product I am now fundraising around. The fact I’m fundraising is a bit of a tell as to where we are really at with solopreneur vibe coding by the way…
- Custom ETL Tools + Business Intelligence: I built (numerous, custom) apps that move data from a source of truth into a data wharehouse that has become a huge unlock for agentic AI to drive insights, intelligence and business decisions that otherwise would go missed. (this is what I do at Guiderail).
- Hitch: A dating utility that helps users identify compatibility before the date. It’s still in production, and people (myself included) use it to save time on bad dates.
- MacOS Video Logger: A native desktop app running a local video-to-text model.
The Hard Truths & The Future
1. “Best Practices” Don’t Exist Yet
We are in the Wild West. It is honestly funny to watch rigid corporate operations try to force their current Agile/Scrum processes on top of agentic workflows. These tools are built for speed and fluidity, not two-week sprints and ticket backlogs. On that note, look at the pace at which Cursor and Replit are shipping using (you guessed it) AI tools and entirely new workflows. Yeah, this is the future, and there are a lot of middle managers and bad PMs in the way right now.
2. The Rise of “Disposable Software”
My gut feeling for 2026? We are going to see a massive rise in Software for One. We will see “turn and burn” software, little programs spun up by AI to accomplish a specific task one time, and then never used again. This requires a radical rethink of what a “program” even is, and yes, I’m aware this is already happening with respect to many of our favorite LLMs.
3. The “Sticky” Problem
Despite the power of these tools, I haven’t seen a fully “vibe-coded” app become a massive commercial success yet. The ceiling right now seems to be “Cool Indie App” that makes $10k-$20k MRR. Why? Because software was never the bottleneck. Distribution and quality were. AI makes writing code easy, but it doesn’t make getting users easy. If anything, distribution will only get harder as the market floods with AI-generated apps.
4. The End of SaaS?
No. But margins will erode. New companies will fill in the gaps that large incumbents ignored because the development cost was too high. Now that the cost of development is nearing zero, those gaps are profitable.
5. So Cursor Wins?
Absolutely not. This battle is maybe just going into the second inning. While I’m a huge fan of the product and am very impressed with the team’s ability to ship meaningful features lightning fast, there is very little lock-in with respect to any of these products. Until someone figures out a way to make switching painful, I wouldn’t bet on any one company winning.
To add, margins aren’t what they used to be in “old” software with new AI tools, with the clear advantage to the foundational model companies here. IE, if cost matters (and it currently does), the model providers win, which gives an edge to Claude Code, Codex, etc. Then again, open source might win out—who knows? We might all end up hosting our own models locally (sounds crazy, but plausible) and just paying a subscription fee to whatever tool interface we like most, rather than a usage fee. No idea.
Final Thoughts
When I step back and look at 2025, I realize I haven’t just been learning new tools; I’ve been unlearning a decade of limitations. For the last twenty years, the defining constraint of the software industry was the “How.” How do I build this? How do I scale this? How do I fix this bug?
Vibe coding is effectively murdering the “How.” – Especially if you think you have an idea the way YouTube, Instagram, TikTok and the iPhone murdered the need for a production studio to go be a video star.
In 2026, the only question that matters is “What.” What is worth building? What problem actually needs solving? What is good and what experience is worth a human’s attention in a sea of infinite digital noise?