Vibe coding is useful for one-off scripts and simple features, while production work needs engineers who can judge the code and guide the agent.
2
Large language models make it easier to reach both good and bad abstractions, so developers should keep repetitive code when that is clearer and sufficient.
3
The people most likely to benefit are experienced developers who can steer agents, provide context, and decide when code is good enough for its job.
Summary
Kitze traces the move from autocomplete and code snippets toward agents that can change whole projects. He argues that the phrase "vibe coding" covers several different practices. Vibe coding can mean accepting generated code with little concern for its structure. Vibe engineering means watching the agent, checking its decisions, supplying detailed context, and stepping in when the work goes wrong. He has used agents to revive side projects, migrate applications, and refactor difficult codebases, but he remains suspicious of generated code because it reflects the limits and mistakes of human-written code. The talk is also a warning about skill and employment. Juniors may struggle to enter the industry if companies replace entry-level work with agents, while experienced engineers who understand systems and maintenance may become more valuable. Kitze is honest that agents can produce useful results quickly, but they can also create a mess that only a skilled engineer can finish.
Frontend development still carries many old problems
Kitze opens by comparing rapid progress in other areas of technology with frontend work. After almost ten years, developers still cannot reliably style every native control, Internet Explorer has only received a logo update in his joke, and the ecosystem keeps debating how to implement a simple counter. React remains the leading library despite its complexity and the fact that developers often disagree about the right way to use it. His point is that frontend development has accumulated a lot of ceremony around tasks that should be simple. That history helps explain why generated code can look acceptable even when experienced developers want to rewrite it.
LLMs make abstraction easier, including the wrong abstraction
Kitze says language models are good at writing React because humans are inconsistent at it. Developers often see generated code and immediately want to abstract it, even when the abstraction adds nothing for users. Agents make this impulse faster. With Composer One, he can reach the right abstraction sooner, but he can also reach the wrong one sooner. He likes that models do not worry about repetitive code. Since 2017, he has argued that developers abstract too early and care too much about repetition. Copying a small piece of code can be a better choice when it keeps the system understandable and does the job.
Vibe coding describes accepting output without caring much about the code
Kitze describes Andre Karpathy's definition of vibe coding as telling the model what to do, pressing accept, and caring less about the resulting code. He connects the idea to an older workflow in which managers assign a feature, test the application, and do not read the implementation. He compares agent coding with a casino: developers buy tokens, press generate, and may receive either a working full-stack app or garbage. The joke works because prompts can consume hours while producing something that could have been written manually in minutes. For Kitze, the term has expanded so widely that people use it for almost any kind of AI-assisted development.
Vibe engineering requires watching and judging the agent
Kitze prefers the term vibe engineering for working with agents while keeping close attention on what they do. He says he has built more than 15 things with LLM and agentic coding that he would not have attempted otherwise. He still watches the code with suspicion because the models learned from human code and human knowledge. His workflow includes examining the UI, describing what he sees aloud, identifying bugs, then moving into the code and explaining what the implementation changed. He also recommends rules, documentation, commands, memories, and a strong starting point of primitives, components, functions, or patterns. The agent cannot hold the entire application in context or read the developer's mind.
Experienced developers should guide agents instead of handing them to juniors
Kitze warns against giving AI tools to interns and junior developers while assuming the tools make them productive. A junior may accept code without knowing whether it is correct. He thinks a skeptical senior developer who learns vibe engineering can produce much better results, although convincing that person to try it is difficult. He places one-off scripts, personal tools, simple features, and code that will not be revisited on the safe side of the spectrum for casual vibe coding. Deciding which code is good enough is itself a skill. It depends on experience with code that does not need further optimization because it already works for the task and its users.
Model choice and speed can change the whole workflow
Kitze says bad experiences with agent coding can come from timing, model changes, too many choices, or using an inexpensive model and expecting the same result. He describes a crowded field in which the answer to "which model is best" changes throughout the day. Composer One changed his workflow because it responds quickly enough for him to watch the agent and interrupt it. Earlier, he would wait so long for a model to finish that he watched videos instead. Faster feedback made coding feel active again, but Kitze says this only helps an engineer who can tell whether the output is right. A fast wrong answer is still not useful to someone who cannot evaluate it.
Good enough code must remain workable for future agents
Kitze says clean code is taking on a different meaning in an agent-driven workflow. It does not need every possible optimization, but it must be clear enough for agents to keep working on it. If a developer accepts sloppily generated code repeatedly, the project eventually reaches a point where the agent cannot move forward. His process is to generate something, inspect the code, briefly test the function, decide whether it is sufficient, and move on. He still recommends studying computer science. In his view, this is a good time to learn because technical knowledge helps people steer models and judge their output.
Agents may remove entry-level work before they remove senior maintenance work
Kitze expects companies to connect agents to the tools developers already use, including GitHub, Slack, Linear, and Sentry. A task that once went to a colleague may instead be assigned to an agent. Over time, several roles around a project could become agent-driven. He thinks the effect may first be felt at the bottom of the profession, where juniors and interns lose opportunities to enter the industry. At the same time, people who maintain old systems have historically been highly paid because the work requires deep experience. He jokes about a future market for engineers who repair the last 20 percent of AI-generated applications and maintain legacy code.
"If you're a vibe coder, you have no idea whether the model is right or wrong."14:56
Who should watch
You are an experienced frontend developer deciding whether agentic coding belongs in production work.
You have tried AI coding tools and got either unusable output or an application that works until the final part.
You are a junior developer, manager, or team lead thinking about how agents will change entry-level work and code review.
