Workshop summary
What is vibe coding
Vibe coding means building software that actually works just by saying what you want in everyday language. An AI agent, a program that can carry out tasks on its own, writes and runs the code. You never have to type code at all, you just describe the result you want.
The term itself is brand new. It first appeared in early 2025, February 2025 to be exact, when Andrej Karpathy, a research scientist on OpenAI's founding team, described on Twitter how he codes now. He hands all the details over to the agent: "give in to the vibes, embrace exponentials, and forget that the code even exists." In short, a new way of making software: the human directs, the AI agent executes.
Why learn it
The way software gets made is being completely overhauled right now. We no longer need to write code by hand. Many companies have already started to stop coding manually and rely on AI agents for everything.
Working this way, you start moving about 10x faster. A prototype that used to take weeks now takes shape in hours. Work that normally takes a month can be done in under a week. A project that used to need a team of 10 can be handled by 2-3 people, and the quality of the software actually goes up, not down.
And this will not stop at software. The same AI agents work for research, planning, reports, analysis, and monitoring, so the same overhaul is coming to every industry. The bottom line: AI agents are hugely influential, and very much worth learning.
How fast the models improved
The pace of change is easier to see for yourself than to take on faith. A research group, METR (metr.org), measures it concretely: how long a task an AI model can complete with about a 50 percent success rate. That number has been doubling roughly every six months.
To make that concrete, here is the progression. GPT-2 could only handle tasks of about 3 seconds, the level of answering a single quick question. GPT-3.5, the model behind the first version of ChatGPT, reached about 36 seconds. GPT-4 reached about 4 minutes, enough to write a whole file of code.
Then close to the start of 2025, o3 broke through to about 2 hours. That is roughly the level of building a simple to-do list app that actually works, and this is when truly modern AI agents appeared, the same moment Karpathy named vibe coding. Today GPT-5.2 sits at about 5.9 hours, roughly a multi-user to-do list with login and stored data. GPT-5.6 Sol is projected to reach about 16 hours in 2026, a full-featured to-do list with plenty of extras. The growth is still exponential, with no sign of slowing down.
You do not need to know code to start
To start vibe coding, you honestly do not need to understand code. In its purest form, you write no code at all and rely on the AI agent for every decision. For now, though, there is a limit, because agents still make mistakes and still depend on you to give the right instructions. It still feels like talking to a genie that does not quite understand what you actually want.
So with no coding knowledge at all, projects stay on the smaller side, roughly the scale of a prototype that takes 1-3 weeks. A landing page or an e-commerce store is well covered. A software product, where the product really is the software, can work too, but keep it simple.
Beyond that, you start to need software engineering knowledge. Software has to be maintained and will keep changing, so you have to explain precisely what needs to change so the agent does not change the wrong thing. The longer a project runs, the more carefully you have to make sure security, performance, and privacy all stay intact. And once your users really grow, you will definitely need to understand code more, because scaling up demands it.
At this level you still have to get into the program yourself, but you rely on AI agents to move much faster: you treat the agent like a very smart coworker, delegate work to it, and still understand and review what it produces. People have started calling this agentic engineering: the engineering is done through AI agents instead of by hand.
What an AI agent actually is
An AI agent is an AI system that works on its own to finish a task. It runs in a loop: act, evaluate the result, adapt, try again, repeating until the task is done. To do that it uses tools such as a browser, a terminal (a text-based way of controlling a computer), and your files.
This is what separates an agent from ChatGPT and other chatbots. A chatbot is just a chat exchange: you send a message, it replies, and that is it. An agent keeps working on its own after you give it an instruction, and it can essentially do anything a human can do on a computer: edit documents, analyze data, do research. Claude and Codex are both agents in this sense.