AI is transforming how software is built. Thanks to vibe coding, almost anyone can create an application without knowing how to program. In the world of business, however, speed alone isn’t enough – security, predictability, and the ability to keep developing the system matter just as much.
Vibe coding is an approach to building software based on “vibe” – that is, on intuition, instinct, and overall impression. In practice, it means that while talking to an agent powered by an LLM, we describe what the solution should look like and what features it should offer, and then iteratively develop the application based on the impression it makes on the user.
This process largely – and sometimes entirely – skips the technical dimension. Vibe coding lets virtually anyone start building applications without having to learn programming, frameworks, or even specific languages. It’s an extremely tempting prospect. We can simply sit down and turn our idea into a working application without spending months acquiring technical knowledge.
What’s more, thanks to how easy it is to work with LLMs, an application can be developed, refined, and expanded in a very short time.
That said, this doesn’t mean vibe coding is without its flaws. While there’s no great problem when our goal is to build an app for analyzing our own workouts or managing a household budget, when it comes to business solutions, “vibe” alone can turn out to be not enough.
The arrival of LLMs triggered a genuine flood of applications offering all kinds of functionality. Anyone with an idea and a bit of determination built their own app for managing tasks, tracking activity, or aggregating news from their favorite sources. Not long after, however, came a wave of attacks on these kinds of solutions. Credit card data, access tokens, and even database passwords were extracted with little effort. The more popular an application became, the faster it turned out that its security policy resembled pinning a list of user passwords to a fence rather than professional protection.
One of the more high-profile examples was Moltbook, a platform described as “Reddit for LLMs.” Its creator emphasized that he hadn’t written a single line of code while building the service. It quickly became clear, however, that the level of security was dramatically low. Using prompt injection attacks, it was possible to remotely execute code on the servers hosting the service. The database also turned out to be improperly secured. As a result, posts promoting cryptocurrencies began appearing at the top of the most-popular rankings, artificially “boosted” to millions of votes. People wanting to promote their projects simply ran commands to increase the vote count above the current record. After a few weeks of operation and patching of the previously discovered bugs, another vulnerability was found – one that allowed full access to all of the platform’s data. Although the service claimed to have 1.5 million active agents, data analysis revealed that only 17,000 accounts were responsible for registering them.
Opinions about the platform were sharply divided – sometimes even within the same person. Andrej Karpathy, a former OpenAI employee and one of the most recognizable figures in the AI world, initially called the platform “one of the most incredible sci-fi takeoff-adjacent things,” only to later describe it simply as a “dumpster fire.” Despite all these problems, the idea was successfully monetized. After less than two months of operation, the platform was acquired by Meta for an undisclosed sum.
Building an application isn’t just about the idea and the ability to implement a specific feature. Knowledge of the technology – its strengths and weaknesses, an understanding of potential attack vectors and protective mechanisms – is also crucial. There are entire classes of vulnerabilities that exploit the characteristic mistakes LLMs make when generating code. Interestingly, models tend to reproduce specific patterns of errors. Paradoxically, such formulaic mistakes can be less dangerous, because they are easier to detect than the unusual vulnerabilities that appear only sporadically.
Among developers there’s even a tongue-in-cheek answer to the security problems of applications built with vibe coding:
„You have to end the prompt with: Make it secure and no mistakes.”
We should all be aware, however, that application security doesn’t come down to a single sentence tacked onto the end of a prompt.
Complete freedom in building applications also has its limitations. There are many ways to solve the same problem, but not all of them are equally good. When building a solution, we give the agent a description of what we want to achieve at a given stage. If the prompt isn’t precise enough, the agent will choose the solution that best fits the current requirements. It may turn out, however, that the next requirement completely changes the assumptions and forces a rebuild of the architecture. Often this won’t be a big problem – it will only mean spending additional tokens to rewrite part of the solution. Sometimes, though, data migration turns out to be impossible, data isn’t saved correctly, or its format makes integration with other systems unfeasible.
Most of these problems can still be solved within the vibe coding approach. The question, however, is whether we want to grapple with such challenges in the case of business systems. Business solutions must be predictable, reliable, and trustworthy. They should have a clearly defined way of operating. In such cases, complete flexibility rarely proves more important than the stability of a solution on which many people’s work and the continuity of an organization depend.
At Productive24AI, we aim to combine the ease of turning ideas into applications – characteristic of vibe coding – with the reliability of solutions built by experienced developers. The agent is given a working environment embedded within a low-code platform. It can operate on configuration artifacts stored in a Git repository and use a set of MCP tools that support all of the platform’s functions. All of the agent’s work takes place within a system that has been established for years and is proven and secure. Every change and operation is validated and verified by Productive24AI’s mechanisms. The knowledge, documentation, and rules for building applications within the platform have been captured as instructions for the agent, while operating procedures have been implemented as MCP tools.
Get to know Productive24AI! Contact us and we’ll be glad to show you what the system can do.
If we ask the agent to build an application for task management, time tracking, or a recruitment module, it will decompose the problem into elements available within the Productive24AI platform. The user’s role remains to properly define the business requirements. The platform’s rules ensure that the resulting solution is ready for deployment and use in a production environment.
The Productive24AI system offers integrations, data import and export, access via REST API, and diverse ways of presenting data – from lists and calendars, through charts, to boards. In addition, it provides a built-in notification system, a messenger, widgets, SSO login, and integration with Extentum AI agents. This brings together all the capabilities of a mature platform used by large organizations with an LLM agent that makes it possible to build ready-to-use applications in a matter of hours — all within a single, coherent environment. You deploy the platform once and then develop it as your business needs change. This is precisely our answer to vibe coding. We call it vibe config.
Vibe – because it offers enormous flexibility in building and lets you focus on the business problem rather than the technical aspects of implementation.
Config – because the result of the work is a configuration: predictable, deterministic, and fully validatable. One that doesn’t allow you to step outside the safe boundaries of the system, guaranteeing the reliability and security of the solution.
Artificial intelligence is undoubtedly changing the way software is built. Thanks to approaches like vibe coding, the process of building applications is becoming faster and more accessible than ever before. But when it comes to business solutions, speed alone isn’t enough. Security, predictability, the ability to keep developing the system, and reliable operation are just as crucial.
The future of application development, then, isn’t about choosing between AI and the traditional approach to development. It’s about combining the flexibility and speed offered by LLMs with experience, proven mechanisms, and a stable technology platform. This is exactly the idea that the vibe config approach delivers – enabling organizations to build modern business applications using AI, without compromising on quality or security.
An MVP in just a few days!
Put Productive24AI’s capabilities to the test in your own processes!
to our newsletter and stay up to date with new publications on the blog.
Try it out at no risk!
