As agents expanded the capabilities of generative AI, from writing lines of code to creating an entire application, industry debate that began with ChatGPT gained new momentum.
term Vibe coding In February, I entered the popular Lexicon from my previous Twitter X post. It was written by Andrej Karpathy, founder of Eureka Labs, an AI Education Company. He is a former director of AI at Tesla and is part of Openai's founding team. In the X-post, which currently has 5 million views, Karpathy described coding as an “approach that completely succumbs to the atmosphere, accepts the exponent, and forgets that codes exist even exist.” He admitted that he rarely checked the work of the tools he was using at the code level. “The code grows beyond my usual understanding,” he said.
To support the generation and modification of Multi-File code, Karpathy coding the atmosphere with a Cursor AI tool called Composer, an extension of the company's original AI-driven integrated development environment (IDE). In April, coding tools like Jetbrain and Github Copilot introduced coding agents, and Windsurf, another AI-powered app generator, was acquired by Openai for $3 billion in May.
Karpathy's original post described coding atmospheric coding as “not too bad for a disposable weekend project,” but some people in the tech industry are taking it more seriously, especially in the startup world. They resumed the refrain, which began with the first introduction of ChatGpt. Developer jobs are about to become obsolete.
There is a new type of coding called “vibe coding.” Here we succumb to the vibe completely, accept exponential functions, and even forget that the code exists. This is possible because the LLMS (e.g. cursor composer w sonnet) is getting too good. Also, just talk to the composer at SuperWhisper…
In a video panel discussion for Y Combinator Leader in March, YC managing partner Jared Friedman said a quarter of respondents in the startup accelerator survey said that over 95% of the codebase was generated by AI.
“This isn't trendy. This won't go away. This is actually the dominant way to code. If you haven't done it, you might just get left behind,” said Garry Tan, president and CEO of Y Combinator, in the video.
Executives from other startups have also purchased the idea.
“I have eight computer screens in front of me, five of which actually do coding the atmosphere, so I'm actively writing code while we talk,” said Ken Schirrmacher, CTO at Oppligeai, a San Francisco passwordless authentication startup, during an interview with Infora Techtarget in April.
There's no longer a typical day for a developer sitting there and writing 30 or 40,000 lines of code.
Ken SchirrmacherCTO, obligeai
“It's something that can be created in a number of teams, a huge number of people, and years, in weeks and months,” Sharmacher said. “The typical days of developers sitting there are writing 30 or 40,000 lines of code solo… You're not starting with a blank slate anymore.”
As atmospheric coding grew, so did dark predictions about the future of human developer work. Dario Amodei, co-founder and CEO of humanity, said in an interview in March that AI will write 90% of the code within three to six months and “essentially all the code” within next year.
“We'll eventually get to the point where AIS can do everything a human can do, and I think that's going to happen in all industries,” Amodei said. “In reality, I think it's better to happen to all of us than it happens — you know, it's a random choice of people. In fact, the most socially divisive outcome is because when 50% of the work is suddenly done by AI, the social message is that we're picking up half of the time.
It seemed that Amodei's predictions could have been realized. Some web companies, such as Shopify and Fiverr, began warning in April that they must prove the value of future human recruitment for AI tools.
Schirrmacher took this one step further. If AI makes software development faster, and there are fewer developers, he predicted that large tech companies employing thousands of developers will also be unsustainable over the next decade.
“Imagine a country that announced that they had scanned a mountain and realized that the entire mountain was filled with gold. The price of gold will plummet,” he said. “That's what AI did in development. And we can't sell products for as long as you're charging today, and we hope to get away from it in 10 years.”
Not so fast: coding agents hit snags
Amodei's prediction that five months later, most code will be generated by AI within 3-6 months, is not near reality, even among large tech companies. In April, Microsoft CEO Satya Nadella estimated that 20% to 30% of the company's code were generated by AI, similar to the estimates made by Google CEO Sundar Pichai in its October 2024 revenue call. In an interview in March 2025, IBM CEO Arvind Krishna predicted that this is where AI-generated code percentages will be maintained.
AI coding takeovers are slower, including hallucinations and ongoing issues with bias or inaccurate outcomes. For most companies with a reputational reputation, Vibe Coding serves as a good way to develop minimally produced products, according to Diego Lo Guidice, an analyst at Forrester Research.
“It will definitely boost the prototyping and innovation of the company, but someone will have to do the rest of their work to turn it into a real, scalable product,” Lo Guidice said. “They can still use code generators, but that is going to be done by a true developer, not a vibe coder.”
This pattern has already begun to change the roles of senior developers in many organizations, but it's not necessarily better.
Kyler Middleton
“We're committed to providing a range of software engineers at Healthcare Tech Company Veradigm,” said Kyler Middleton, leading software engineer. “Personally, I'm reviewing 10 times the amount of code I last year, but I can't run it at a 10 times faster pace to ensure that all the code generated by AIS meets the criteria.”
Over time, junior developers are likely less skilled at writing their own code, increasing the skill gap between junior and senior engineers and relying on atmospheric coding tools, according to Rob Strechay, an analyst at TheCube Research.
“I think it hurts young developers so badly and affects the way companies 'grow' their developers,” Strechay said.
He said that continued macroeconomic volatility and uncertainty are only making it even worse.
“The team is not receiving the employment they thought they would do and instead they are told to use AI to be more productive,” says Strechay. “At the same time, they're slower when using AI on potential IP issues. AI might be as good as university graduates, but using AI takes a lot of time to “fix” the code.
Google's Dora has already observed this slowdown in research over the past year, noting in an April 2025 report that AI assist code has not committed to developers in any other way.
“AI is creating tasks that people think are valuable faster and more valuable, but it's not really helping tasks that people don't enjoy,” the report reads. “This is what's happening while the struggle and burnout remain unchanged and stubborn in the face of AI adoption, highlighting that it hasn't cracked the norms that AI can help avoid meetings, bureaucracy and many other hard work.”
Long view: The decline of development and democratization
However, some believe that, as AI technology solves the initial problem, developer jobs will disappear for years rather than months.
For example, AI code can introduce security vulnerabilities, and it's difficult for humans to keep up with volumes, but “but the solution is to pack more AI and AI validate the code,” according to Schirrmacher.
Some high-tech vendors believe that the long-term outcome of this cycle of managing AI using AI is the complete autonomy of future agent AI systems and the removal of most human engineers.
fetch.ai is developing Agent AI PAAS based on its experience as a cryptocurrency company and will use blockchain consensus mechanisms to verify the accuracy of agent-based output. Founder and CEO Humayun Sheikh predicts advances in quantum computing and other areas will accelerate autonomous AI systems.
“We need to build this new agent web and people need to build that infrastructure,” Sheikh said in a February interview. “But once it's done, you don't need to write the same code in 10 different companies, 10 different companies. Software development is a thing of the past, as LLMS can do a better job.”
However, there have been few major changes in business technology, such as the transition from mainframes to distributed systems or from on-premises data centers to cloud computing. With the rise of DevOps, similar predictions about the end of IT operations experts have not been realized. Instead, roles have evolved into platform and site reliability engineering.
For now, most industry experts have predicted a similar trajectory for developer employment in the age of AI coding.
“We're committed to providing a range of services to our customers,” said Nick Cassidy, lead innovation engineer and lead AI product developer at Stellarus Group, a San Diego health services company. In an interview at the Red Hat Summit in May, Cassidy stressed that his personal opinions do not necessarily reflect the opinions of his employers.
“The development openings are expanding. You may not need to have the deep skills you need to get started as a developer,” Cassidy said. Tools such as Red Hat's Instrublab have enabled citizen developers to build AI-driven applications and fine-tuning models, he said.
This led to new ways of working in some companies, and business leaders have used vibe coding to speed up handoffs to engineering teams.
“I've personally been taking AI tools to prototype things for engineers,” says David Strauss, CTO at WebOps service provider Pantheon.io. “It's much easier to share partially wrong prototypes than natural language.
Generation AI also generated new areas, such as rapid engineering, which could indicate the future of software development work. Even if daily development tasks change, the key needs of human problem solving remain, predicted Lars Maaløe, co-founder and CTO of CORTI, a healthcare AI company.
“AI models tend to regress towards the average,” Maaløe said. “So what they know, they're a very comfortable building. So, if you ask them to build website number 1,000, they'll build website number 1,000 with the design schedule that is usually used to look.
Beth Pariseau, senior news writer at Informa TechTarget, is an award-winning veteran of IT journalism covering Devops. Any hints? Please email her Or reach out @pariseutt.
Kyler Middleton
