Vibe Coding ❌ Agent Coding ✅
Vibe coding is a nowhere going hype, but Agent coding is real and it’s changing the way we build software forever.
The vibe coding hype cycle
LinkedIn, X and Substack are full of people talking about Vibe coding. But what is it and more importantly is it all that it is hyped to be?
The term “vibe coding” was coined by Andrej Karpathy in February 2025 . Karpathy is a computer scientist, co-founder of OpenAI and former AI leader at Tesla .
What Karpathy meant by “vibe coding” was a fundamentally different approach to programming that leverages AI capabilities. He described it as a practice where you “fully give in to the vibes, embrace exponentials, and forget that the code even exists”.
The basic idea is to rely entirely upon LLMs and code by only using natural language prompting instead of writing the code yourself.
Karpathy explained it as “it’s not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works”.
This approach means building software with an LLM without reviewing the code it writes - essentially a “forget that the code even exists” approach.
Instead of methodically writing syntax, developers now describe what they want in natural language, and Karpathy’s concept recognizes how sophisticated AI systems have evolved, with LLMs advancing to a degree that nearly eliminates the use of traditional coding mechanisms.
The term captured the zeitgeist of AI-assisted programming so effectively that it was listed in the Merriam-Webster Dictionary as a “slang & trending” term just one month after being introduced .
Wait a minute!
Forget that the code exists approach? That may be fun for prototyping or researching the LLM’s capabilities. But this will not build you a complex flawless working platform or AI service.
What I do believe and what is already changing the way we build software forever, is Agent coding.
Coding with help of AI where you do acknowledge the code exists.
Where the person instructing the agent is constantly verifying the proposals, solutions and code the agent provides.
Not the verify the code is correct, but to verify if the Agent understood what you were trying to communicate.
Not to verify if the proposed solution is correct, but to verify if the correct conclusion was drawn from the facts laid out, or issue encountered.
To be able to do that, the coding agent you are using needs to provide an extreme high level of transparancy on its actions.
Very much like how a founder or product manager steers its developers. However these are now replaced by the AI agent.
Critical developers
I’ve noticed some developers complaining current coding agents are far far away from being able to build complex code.
They may indeed not yet be where we want them to be, but as much as your development team needs good leadership, so does the coding agent. And if you are good at coding, it doesn’t mean you are a good leader too.
The quality of the output of the coding agent depends on the quality of your input and your skills to lead the coding agent to the desired result.
Developers by nature think in solutions, but leading a team or a coding agent requires you to lead at a meta level and at the same time understand the inner workings of your product (code, infrastructure, functional behavior) in more detail.
That meta level means your decisions must be geared toward realizing the overall vision for your solution (involving business model, market positioning, resource allocation, go-to-market strategy, partnership decisions, product roadmap prioritization, and strategic pivots).
And exactly that maybe hard when you’re used to diving deep into code to bug fix or find solutions, or try to bring a design to life.
That’s why I think agent coding will not benefit developers most, but will first of all benefit founders and product managers that understand how to lead at meta level and are able to dive into the details at the same time.
Coding agents and the way you are now able to design, build, ship and test super faaast have already changed development cycles at early adapters (building AI native software).
And they will increasingly impact the way teams work, and support creators to bring their ideas to life without limits.
What made you an expert, exactly?
Oh - you built an AI travel platform in a weekend?
Yes I did, and I never thought it was possible.
I started building Ask Seve (https://askseve.com) to create a community based travel journal. As a first prototype, it turned into a full blown travel planner based on the same principle.
It took me a weekend to build the base on https://lovable.dev two weeks to make it into a grown up platform with a full working backend and admin console. Burning credits like 🔥🔥🔥. Properly based on product-led go-to-market principles.
With prompting alone. I did not code a single thing.
I was suprised, having tried to implement AI in existing software before, my advised route was a baby step implementation.
And that is still very true. But - now I experienced - only in case your product was built before the AI age.
If you now build AI native, you really can forget all that baby steps talk, and build without the constraints that your current tech brings.
Oh so liberating.
So should you believe the hype cyclers?
Can we vibe our way to code a platform without thinking about code?
No, you can’t. But do believe the capabilities of Agent coding. And if you are a designer or developer start learning to lead a coding agent. If you are a creator start believing you can materialize your dreams.
Is building with AI only for “pet-projects”, or are founder led AI projects a security or future bug fixing nightmare?
That’s just the haters talking.
You do need to be critical in what the AI coding agent serves you, and without a doubt it helped me having over 10 years experience in leading a tech team. And yes there will be testing and bug fixing.
But the development cycles have dropped drastically, it’s craazzzy.
Time to document changes, tasks, development milestones? The update has been implemented before you can say “document”.
And because errors can be fixed quickly pushing code quickly isn’t an issue either.
Not that I advise in a full client environment to push major changes untested, but I found myself pushing not direct client side changes quicky to test in production. Any errors were quickly mitigated.
And the best thing I noticed as a result of these micro development cycles: the serendipity built in.
I can now immediately respond and update direction of where I was headed in response of the Coding agent’s update or reply. That means your project’s road to the desired goal becomes much more fluid and adaptable.
Conclusion
The Vibe Coding Hype Cycle
The term “vibe coding” has sparked intense debate across LinkedIn and developer communities, with many caught up in the hype cycle surrounding this new approach to AI-assisted programming. While Andrej Karpathy’s concept of “forgetting that the code even exists” captures the revolutionary potential of LLMs, the reality is more nuanced than the hype suggests.
Beyond the Hype: The Reality of AI-Assisted Development
The pure “vibe coding” approach—relying entirely on natural language prompting without reviewing generated code—may work for quick prototypes or exploring LLM capabilities, but it falls short for building complex, production-ready platforms. The idea that we can completely abandon understanding our codebase is both naive and potentially dangerous for serious software development.
The Real Revolution: Agent Coding
The true transformation isn’t in “vibe coding” but in agent coding—a more sophisticated approach where developers work collaboratively with AI agents while maintaining oversight and understanding of the code being produced. This method requires:
- Continuous verification of the agent’s understanding of requirements
- Critical evaluation of proposed solutions and architectural decisions
- Transparent communication between developer and AI agent
- Leadership skills to guide the agent effectively toward desired outcomes
A New Skill Set for Developers
The emergence of coding agents demands a fundamental shift in developer skills. Success requires moving beyond pure technical implementation to become effective leaders of AI agents. This involves:
- Meta-level thinking about all founder-level decisions geared toward realizing the overall vision.
- Deep product understanding across code, infrastructure, and functional behavior
- Communication skills to articulate requirements clearly to AI systems
- Quality assurance mindset to validate AI-generated solutions
Interestingly, founders and product managers who already possess these leadership and communication skills may initially benefit more from agent coding than traditional developers who are accustomed to diving deep into implementation details.
The Transformation is Real
Despite the hype, the fundamental capabilities are genuine. Building AI-native applications from scratch allows developers to bypass legacy constraints and achieve dramatically accelerated development cycles. The key is approaching this revolution with both optimism about the possibilities and realism about the requirements for success.
The future belongs not to those who blindly embrace “vibe coding” hype, but to those who thoughtfully integrate AI agents into their development process while maintaining the critical thinking and oversight that complex software demands.
Keep pushing, keep creating,
//Caroline Vrauwdeunt
P.S.
Would you like to get in touch?
You can find me on LinkedIn or Twitter.
If you wish to make a request on particular topic you would like to read, you can send me an email to caroline@andrs.nu .