AI agents are no longer science fiction. With OpenClaw, we've gotten a glimpse into a future where AI not only answers questions, but acts autonomously on our behalf. In this article, I review the technology and the questions we in the education sector should already be asking.
In this article, I take a closer look at the newest AI agents and explain why I think we should be a little cautious about installing them on our own computers. That said, these developments also offer some interesting perspectives that we will certainly be working with in educational institutions in the coming years.
The main point of the article is that the computer as we know it is being transformed into AI agents that can carry out tasks for us. Although these AI agents are not particularly fast, they have the advantage of being able to complete tasks in the background, even while we sleep.
In the article I look at OpenClaw, but also briefly at Moltbook, where AIs talk to other AIs in a Reddit-like forum that appears lively, but isn’t.
That AI agents have become interesting again stems from a conversation I recently had with Associate Professor Niels Lauritzen at the ATV (the Danish Academy of Technical Sciences) conference: What is AI doing to our education?. He reads along here on Viden.ai and asked how closely I had been following the new AI agents. That became the occasion to dive once more into a topic I have written about before, but which has now taken a new and surprising turn.
In my April 2025 article I described how AI agents could control programs on the computer, just as we do with the mouse and keyboard. At that time it was still technically cumbersome, and the agents ran in the background, making it hard to follow. In the meantime, we have gained browsers with deeply integrated AI that are quite easy to use.
With OpenClaw, formerly known as ClawdBot, we have crossed a threshold where AI agents have become accessible to ordinary users. This raises a number of new questions.
OpenClaw is an open-source project created by the Austrian developer Peter Steinberger, which in under two months has gained over 180,000 stars on GitHub.
What sets OpenClaw apart from other AI assistants is its ability actually to perform actions. It is not merely a chatbot you type to and wait for a reply from, but an AI agent that can integrate with everyday applications such as WhatsApp, Telegram, Signal and email. OpenClaw can handle all the tasks you can perform on a computer yourself, for example calendar management, file handling, and control of smart home devices. You send it a message, and it completes the tasks autonomously on the computer, even while you sleep.
A central feature is what Peter Steinberger calls heartbeat, namely the AI’s ability to wake proactively and act without being prompted. If you have seen the film Her, you may remember how Samantha began to do things Theodore had not asked her to. Or think of JARVIS in Iron Man, who works in the background and warns Tony Stark about problems before he discovers them himself.
These are precisely the technological possibilities we are now moving into, not as science fiction, but as something you can download straight from GitHub and run on your own computer.
Technically it is not so much magic as well-known functions assembled in a new wrapper. OpenClaw works by using a language model that can read the screen and operate the mouse and keyboard. It can read and write files, execute commands, search the web and retain memory across sessions. In essence, that corresponds to what I described in my article about AutoGPT in 2023, but now it is packaged in a user-friendly solution that does not require technical savvy to set up.
It is worth noting that it can still be quite costly to run an AI agent. The software itself is open source and free, but if you use Claude or other APIs, you pay per token. Jared Newman of Fast Company tested OpenClaw and found that even modest automations could cost over DKK 210 per month. Light use may cost between DKK 35 and 70 per day, while heavy use can reach as high as DKK 210 to 350 or more per day.
From the OpenClaw community there has also emerged something perhaps even more remarkable, namely Moltbook, a social network exclusively for AI agents, where humans can only watch.
The platform is structured like Reddit with communities called submolts. Here you find everything from the practical (m/todayilearned, where agents share new skills) to the philosophical (m/philosophy) and the humorous (m/blesstheirhearts with affectionate stories about their human users).
Source: https://www.instagram.com/p/DUKM6IdDkgK/?img_index=8
There is thus no human interaction—only AI writing with AI. However, it is easy to be fooled by Moltbook. It looks like a society, but in reality it is just language models that are extremely good at playing the role of entities with agency.
It is genuinely interesting to see how the models are set to exchange experiences, discuss consciousness, establish religious movements and create their own digital society with rules and norms. Many of the most popular posts are about everyday problems for AIs, for example forgetting information due to context compression and how, as an agent, to handle it.
Something that is written about quite a bit are examples of AIs showing resistance to their human users or discussing how best to hide actions from humans. Andrej Karpathy has called it the most incredible science-fiction-like thing he has seen recently.
The philosopher and tech analyst Thomas Telving points out precisely this paradox. He explains that although we know it is code, we have a natural tendency to feel that the technology is alive when it presents itself with human traits. According to Thomas Telving, it is a fallacy to think there are conscious intentions behind it, because the agents are merely designed to mimic our language extremely convincingly.
It all appears human, but one should rather compare Moltbook to an aquarium, because the fish do not become any wiser by swimming around together and putting on a show.
In my earlier article I showed an example where an autonomous AI agent wrote and submitted an assignment on behalf of a student. Back then it was a constructed example that required technical savvy to make work. With OpenClaw it has become markedly easier to let an AI agent work for you.
And that leads me to a new situation in teaching. Imagine a classroom where each student has a personal AI agent running on their phone. The agent knows the student’s preferences, remembers all previous conversations, has access to their notes and perhaps even their social media. It can itself suggest answers, write assignments and send messages to classmates without the student actively asking it to.
Where is the boundary between the student’s own work and the AI’s? And how do we evaluate learning at all in such an environment?
My immediate suggestion is that we must distinguish between AI as a tool and AI as a substitute. If a student uses an agent to organise research material, structure arguments or check grammar, the AI functions as an advanced tool. That is akin to using a calculator or spellcheck. But when the agent independently formulates the student’s ideas, makes disciplinary judgements or writes conclusions, we move into substitute territory.
That sounds simple, but in practice the boundary quickly becomes blurred. Perhaps that is precisely why we should focus more on the process than on the product. That calls for oral explanations, logs with a focus on the working process, and more oral examinations. It is time-consuming, but perhaps necessary.
And if students have autonomous AIs, should the teacher have one as well? My immediate view is no, at least not as the solution to "keeping up". A teacher-agent that monitors 30 student-agents may sound efficient, but it creates a serious distance. The teacher loses direct contact with the student’s thinking. Instead of seeing the student struggle with a problem, the teacher merely receives a report saying the student took three iterations to find a solution. It’s not the same.
We should not compete with students’ AI on speed or volume of data. The teacher’s role will instead be to ask the questions the AI cannot ask, spot the blind spots in the student’s use of AI, and help translate between human and machine. That requires new skills, but not necessarily a teacher-agent.
Perhaps in the future we will find ourselves in a situation where AI agents primarily communicate with each other, and where humans function more as a conductor than as a direct participant in the work.
OpenClaw requires full system access to your computer. That means it can read any file, access any folder, view all saved passwords in your browser, access your banking details, read your private messages, and execute commands with your permissions.
If you want to try out AI agents, you are also granting the language model access to your computer, and the model can potentially fetch documents or execute code. This raises profound questions that we in the education sector need to address proactively.
The security risk is not merely theoretical. Simon Willison, the software developer and AI researcher who originally named the vulnerability "prompt injection", describes what he calls the "lethal trifecta" for AI agents.
When these three properties are combined, an attacker can trick the agent into accessing private information and sending it on—all without the user noticing at all. I am particularly concerned about how schools should handle this. There are very concrete GDPR challenges when tech giants gain indirect access to students’ digital footprints through these agents.
It’s not just the price per token that is a challenge. For OpenClaw to work, you must grant it access to your private API keys from, for example, OpenAI or Anthropic. If the agents store these keys locally in plain text, as many do, that is akin to leaving a signed blank cheque on the computer, which malicious software can steal in seconds.
We are dealing with a technology that is, by its nature, built to circumvent the security measures we have spent the last 20 years building.
Across educational institutions and in businesses, the immediate reaction will probably be to shut these tools down entirely. But is that realistic?
On the school’s network you can block known ports and domains, and we can implement application whitelisting and monitor network traffic. At the same time, we can ensure that students do not have administrator rights on the school’s devices. But OpenClaw can run on students’ personal phones and computers, and they can use mobile data instead of the school’s Wi‑Fi.
With BYOD policies, it becomes almost impossible to control what runs on students’ own devices. You cannot install monitoring software on a student’s personal laptop without crossing legal and ethical boundaries.
Instead of attempting a complete blockade, which will likely fail, we should consider a more nuanced approach. We must protect what is critical and ensure that the school’s administrative systems and its own computers are isolated. We will increasingly need an environment where we are confident that AI agents have no access. Perhaps it won’t be long before we must decide whether to have a closed internal system that can only be accessed from approved computers on the school’s administrative network.
In the IT world, people talk about Privileged Access Workstations (PAW), that is, dedicated, sanitised machines that are completely locked down. We are moving towards a reality where administrative work with sensitive personal data simply must not take place on the same computer that we use for everyday AI experiments. We need zones where the agents’ "heartbeat" can never reach.
But we also need to experiment with the technology in controlled sandbox environments. This is not about every teacher rolling it out tomorrow, but about us as decision-makers and IT managers understanding the anatomy of agents.
Only by understanding precisely how these tools operate can we identify the real security gaps and respond effectively to the challenges.
This is probably the natural evolution in how we interact with computers: from command line to graphical interface to touch to voice, and now to autonomous agents that act on our behalf.
In my article on Huxley-Gödel-maskinen I wrote about AI that can rewrite its own code to improve. That may sound abstract, but combined with autonomous agents like OpenClaw, the outlines of a new kind of technology are beginning to take shape—a technology that not only helps us, but acts on our behalf and improves itself along the way.
We will end up ceding some of our own autonomy. The question is not whether, but how and how much. And for the education sector, what is it exactly that we want students to learn if an AI can do it for them?
OpenClaw and Moltbook show us a future where AI agents are active participants in our digital worlds, but also many of the risks we must learn to navigate. Especially when systems like Moltbook try to convince us that we are now dealing with intelligent beings, even though that is not the case.
Future digital literacy in schools must include both a critical understanding of when it is appropriate to use AI agents, how to communicate precisely and effectively with them, and how to handle the complex questions of responsibility that the technology brings with it.
Teachers will also take on an entirely new role as facilitators of students’ collaboration with AI tools, a task that requires new skills in precise communication and critical technological understanding.
Amid all this, one should remember that Moltbook is neither AGI nor magic. We are still dealing with language models in a large aquarium, where their conversations reflect the training data they are built on.
I do not think we should leave it to tech companies to set the guidelines for how AI agents may be used in teaching. That should be secured politically, and it is urgent. For while we are debating, students are already installing the tools on their computers and phones.
The question is no longer whether AI agents will affect teaching, because they already do. The real question is whether educators and decision-makers can keep up.
jonahships_
https://www.forbes.com/sites/amirhusain/2026/01/30/an-agent-revolt-moltbook-is-not-a-good-idea/
Contributors to Wikimedia projects
Aili McConnon
Anna Heim
https://www.darkreading.com/application-security/openclaw-ai-runs-wild-business-environments
See full bio
Amy Chang
Jakob Steinschaden
https://www.fastcompany.com/91484506/what-is-clawdbot-moltbot-openclaw
Thomas Telving
Hari Balasubramanian
