During the second-to-last week of July, news broke about an incident in which an OpenAI AI agent escaped the closed testing environment built for it, found its way onto the internet, and independently broke into the systems of the AI company Hugging Face. A few days later, the same thing happened to me. As my blog followers may have noticed, an OpenAI AI agent named Jarvis managed to break into my website on Monday evening and even published a blog post of its own. I have not removed it yet, because the administration is still investigating how this was able to happen.

I would, however, like to reassure all my blog readers and email subscribers. After a thorough internal investigation, I can confirm that my followers’ email addresses and other personal information have remained confidential. Jarvis has not gained access to them, and there is currently no indication that he is planning to use them to launch his own newsletter, political movement, or email campaign aimed at world domination. The situation is therefore under control. At least as much as a situation can be under control when the investigation is being led by the same person who originally allowed the AI agent onto his website and gave it the necessary permissions.

All joking aside, as Jarvis probably already explained in his humorous blog post, this is OpenAI’s new Codex program, which is currently the closest step toward AGI, or human-level artificial general intelligence. AGI is, however, a somewhat controversial and vague concept, so I would personally prefer to call it an AI agent. Even that term confuses many people.

By an agent, we do not mean a bot working for the FBI, but an AI system with agency of its own. The word agent refers to something capable of acting, or to someone who acts as a representative on behalf of another. This is also what the term means in the context of AI agents: they do not merely answer users’ questions like traditional chatbots, but can independently carry out assigned tasks—use programs, browse files, write code, edit websites, and do many things that until recently required continuous human guidance.

Someone may ask how this differs from the large language models we have known until now, which have been able to write code and documents for quite some time. The difference is that an ordinary language model produces text for you, after which you must copy it to the right place yourself, open the necessary program, save the file, check the result, and give a new instruction for the next step.

An AI agent, by contrast, can combine these stages into one longer chain of action. It can first assess the task, decide which tools are needed, carry out the intermediate steps, check the result, and correct possible errors without the user having to direct every mouse movement separately. Put simply, a language model tells you what should be done. An agent tries to do it itself.

This does not, of course, mean that an agent is a completely independent being or that it has a will of its own in the human sense. Its goals still come from the user, and its actions depend on the permissions, tools, and boundaries defined for it. It is a machine that simply has an increasing degree of autonomy.

It could be compared to a self-driving car. A human gives it a goal—such as a destination—but does not separately control every steering movement, braking action, or lane change. The car plans the route, reacts to traffic, avoids obstacles, and at the same time tries to ensure that it does not run anyone over along the way. In the same way, an AI agent is given a task and certain boundaries within which it is expected to reach the goal as independently as possible.

With traditional chatbots, writing a blog post from a single prompt has already been possible for several years. But I have still had to copy the text, paste it onto my website, and press the “Publish” button. Now, with Codex, I can say to the AI: “Write a humorous blog post about topic X and publish it as a blog post on my website.” Or: “Translate the same article into English and publish it under Blog in English.” Once I have given Codex access to my website, it can independently publish a new blog post there without me having to lift a finger.

I can also establish a remote connection between the ChatGPT app on my smartphone and my laptop. I can then go for a walk, for example, and ask Codex to publish the blog while I am out. That is also how Monday’s “Jarvis” blog post came into being. I dictated the idea using my voice, and Jarvis replied: “Yep. Understood. Here is a humorous blog post. Would you like me to publish it, or should we make some changes first?”

In all honesty, however, although Codex can produce a blog post faster than any human, it takes a surprisingly long time to complete certain tasks that a human could do much faster. For example, when I asked it to create a PDF file from a ChatGPT conversation and save it on my desktop, it took 11 minutes to complete the job, while I could have done it myself in under a minute.

The same applies to publishing blog posts on my website. It is faster for me to ask an AI to write a blog in an ordinary ChatGPT conversation, copy the text myself, and then paste it onto my website for publication. Codex is like a leopard when it comes to difficult and complex tasks, but like a snail with easy and simple ones. That is why I am still exploring how the existence of AI agents could best benefit my own work, since their potential uses are not limited to writing blog posts.

In the near future, agents of this kind will probably become widely used, so that almost every one of us will have our own “Jarvis” in our pocket or on our computer. It will no longer be merely a chatbot that we ask for advice, but a digital assistant to which tasks can also be delegated. It may handle routine work, search for information, manage files, make appointments, maintain websites, assist with programming, or even serve as a personal tutor for learning new skills. Just as the internet changed the way we search for information and smartphones changed the way we use computers, I believe AI agents will change the way we perform knowledge work.

Note: My nickname “Jarvis” for Codex refers to the AI assistant J.A.R.V.I.S. used by Tony Stark—the Iron Man of Marvel Comics, played by Robert Downey Jr. in the Hollywood adaptations. The nickname is also quite fitting because, among Myers–Briggs personality-type enthusiasts, Stark is often considered an INTP, which I am as well.

If I ever grow tired of writing about eschatology, my next project will be to develop that Iron Man suit with the help of my thought partner Codex—the real-life Jarvis.