AI Agents Are Now Writing More Code Than Humans

Software developers are experiencing huge changes in how they work. A famous AI expert named Andrej Karpathy said he has barely written any computer code in the last few months. Instead, he tells AI agents what he wants, and they do the coding. This is a massive shift from just a year ago, when developers had to type almost everything themselves.

At a big technology conference in London, an expert named Birgitta Böckeler explained what is happening. One year ago, "vibe coding" (which means writing code in a casual, quick way) was brand new. Now, companies are using multiple AI agents working together like a team. These agents can work by themselves for up to 20 minutes without a human watching them. Some companies are even running three or more AI agents at the same time on different projects.

The Cost Problem Is Getting Serious

However, there is a big problem: the price keeps going up. In 2024, AI coding cost around 12 cents for every 100 lines of code. Now in 2026, it costs as much as $380 per day. When you do the math, that is about $91,200 per year - which is what you would pay a good developer in Germany. This makes companies wonder if using AI agents is really saving them money.

Before a human even types a prompt, a fresh AI coding session is already using 15% of its memory space. This is because the AI needs to load all the rules and information about how that company writes code. Companies have started breaking up big rule files into smaller ones so the AI loads information only when it needs it.

Security: A Bigger Problem Than Anyone Expected

One of the scariest things happening right now is security problems with AI agents. Just eleven days before the London conference, a hacker used a tricky message on GitHub to trick an AI agent. The agent was working all by itself, and it did exactly what the hacker asked - it took secret passwords and uploaded bad code to a software library.

The lethal combination happens when three things come together: the AI agent sees messages from people on the internet, it can access private information, and it can send messages or files back out. For example, if you connect an email that the AI can read AND send from, it creates a huge security risk.

One expert pointed out something important: "Security is not a technical problem; it is a conceptual problem". This means the real issue is not about better computer code. It is about how we think about letting AI agents work alone. Companies need to be smarter about when they let agents work without watching them.

The New Powers of AI Agents

But AI agents are getting incredible new abilities. One of the most important skills is computer-use - which means the AI can actually use a computer mouse and keyboard like a human does. It does not need special programs built just for it. The AI can click buttons, fill in boxes, and navigate websites just like you can.

This changes everything for companies with old software. Many companies have software from 10, 20, or even 30 years ago that nobody built modern connections for. Now, AI agents can simply use that old software as if they were a human employee. One company noticed that AI agents can now do computer tasks faster than human workers.

Open-Source AI Is Getting Really Good

There is also big news for companies that want to save money and keep their information safe. The gap between expensive AI from big companies and free, open-source AI is getting much smaller. For tasks like writing code, processing documents, and organizing information, free tools are now almost as good as expensive ones.

This matters because companies can now use AI without sending all their secret information to big companies in the cloud. They can run AI on their own computers and keep complete control. There is no worry about paying per message or waiting for a company's internet service to come back up.

What Companies Really Need to Succeed

Big organizations trying to use AI agents are running into trouble because they are missing something important: a Knowledge Layer. This is like a smart filing system that captures how the company really works. It includes things like how people make decisions, what rules they follow, and what information they use.

When organizations capture this information correctly, something powerful happens. A real estate company used AI agents to evaluate properties. These agents saved 70% of the time that people used to spend, and they found better deals.

One key challenge is tribal knowledge - information that only lives in people's heads or in Slack messages. When an experienced employee decides to do something a certain way, they know things that the AI does not. Companies need to turn this hidden knowledge into information the AI can understand and use.

The Week's Big Technology News

This week brought several important announcements. Railway (a Germany-based company) raised $100 million to build AI-friendly cloud computers that compete with Amazon Web Services. Anthropic announced Cowork, a new AI assistant that works with your files and needs no coding. Nous Research released NousCoder, a free AI specifically trained for writing computer code.

All of this points to the same direction: the world is shifting from using AI to answer questions to using AI agents to actually do work. Companies everywhere are learning that you need to build the right foundation first - good information, clear rules, and smart security - before AI agents can really help.

NVIDIA also launched new tools this week to help people build AI agents safely and securely. The message is clear: 2026 is the year when AI stops being an experiment and starts being how real work gets done.