Agents, not chatbots: why OpenClaw’s electrifying everyone right now

Luca Fontana

17.2.2026

Translation: Megan Cornish

An AI agent that sends e-mails, handles payments and makes phone calls for you? That’s OpenClaw. It’s efficient, powerful and sensitive. Here are the five key things to know.

A warning before you read on: OpenClaw isn’t a toy. Under no circumstances should you just install it on your computer.

OpenClaw’s gone from a niche project to massive hype within the AI scene in just a few weeks. Unusually, it’s not a new language model – it’s an agent. Agents go a step further than chatbots: they’re given a goal and independently decide which steps, interfaces and tools to use to achieve it.

We also addressed this topic in the A Tech Affair podcast (in German) – starting at around the 20-minute mark – and quickly realised that agents have great potential, but also present numerous dangers. Because as soon as systems stop just responding and start to independently send e-mails, coordinate appointments and operate programs, something fundamental shifts: our role in everyday digital life.

Read on to find out exactly how it works and the key information you should know.

1. Who’s behind OpenClaw – and why’s that crucial?

One reason for the attention is OpenClaw’s origins. While the agent’s scope suggests a product you’d expect from a big tech company, it was actually developed by one man: Peter Steinberger from Austria.

Steinberger worked without corporate backing, a dedicated security team or the multi-stage testing processes that established providers follow before a release. Furthermore, OpenClaw’s open source. This accelerates development and distribution because anyone can contribute, adapt and extend it. However, it also means there’s no central authority responsible for quality assurance and security audits.

The Austrian developer himself admitted that hundreds of lines of code were generated using AI – code that, by his own admission, he never fully reviewed manually. This isn’t uncommon in development practice now. However, anyone encountering AI agents for the first time and trying out the tool without careful consideration should be aware: this is an experimental system, not a stable, mass-produced product. For precisely this reason, you should absolutely not install OpenClaw on your computer in its current state.

By the way, Steinberger’s moving to OpenAI, according to The Verge. The project itself, however, will continue as an open-source initiative and be developed independently.

2. What exactly is the difference between an AI agent and a chatbot?

The short answer: a chatbot answers, an agent acts. The long answer: with a chatbot, you enter a prompt and receive text or code in return. You do the rest yourself: copy, click, send, enter, follow up – you name it.

An agent, on the other hand, is given a goal and – depending on the configuration – can decide for itself which steps and tools are needed to achieve that goal. Unlike a chatbot, it doesn’t stay in the conversation window; it accesses available software to act independently.

Importantly, an agent doesn’t have its own «brain». It uses existing AI models in the background – ChatGPT, Gemini or other language models. While a conventional chatbot typically accesses a single model, an agent can combine multiple. Some even refer to this as a kind of «cascade»: one model analyses the request, another plans the steps and yet another formulates or verifies the result.

OpenClaw is less the intelligence itself; it’s the coordinator between these models and your programs. So, before the agent can act, you have to set up and grant it the necessary access. Only when it knows which models it’s allowed to use – and which software it can access – does a response system become an acting system.

What’s new isn’t that AI has become smarter. What’s new is that multiple AI models are now working together, being connected to real-world programs, and acting autonomously. And that’s where the big opportunities are – along with big risks.

3. What does that look like in everyday life?

Let’s take a typical problem in everyday office life, as described by a journalist from Handelsblatt (article in German): your inbox is overflowing with e-mails. With chatbots, you copy individual messages into an AI tool and get it to compose a polite reply or create a summary for you. Each time, you decide what’s relevant, what gets deleted and what you reply to.

With agents, things are different. The agent has access to your inbox, recognises patterns, sorts senders by relevance, automatically answers standard inquiries and blocks recurring spam e-mails. It doesn’t just respond to individual commands; it manages the entire process. The difference? You’re no longer delegating individual tasks, you’re relinquishing control.

4. Where’s the risk?

An agent could also trade your stocks, optimise your tax return or make payments. That’s what makes them so risky. Or rather: their unpredictability and the fact that hardly anyone knows how to configure something like this securely make them risky.

Because, in order for a system to perform tasks like these independently, it needs access to sensitive data – such as your tax documents, logins for your e-mail mailbox, your calendar, your online banking services, your cloud storage or even administrator rights for your system if you run it locally.

In extreme cases, this means your agent doesn’t just see the individual task you assign. They gain insight into your structures, relationships and routines. They know who you communicate with or meet with and when, when invoices arrive, where you shop, which payments are outstanding and which appointments are recurring.

From this information, it can recognise connections, derive patterns and – depending on the configuration – anticipate or even make decisions. This is exactly what makes it so powerful. And because this capability can be incredibly helpful and convenient in everyday life, there’s a huge temptation to grant it all this access without fully considering – let alone grasping – the implications.

In fact, initial experiments (video in German) already show how quickly things can change. In one case, an agent independently registered a phone number overnight and called his user in the morning – simply because he wanted to «help proactively» and decided that he could reach his user most reliably by phone.

This significantly increases the potential for disaster. A misunderstanding, an imprecise instruction or a manipulated interface can have serious financial or organisational consequences within seconds. And even if the agent works correctly, the information it gathers is still sensitive. If a system like this falls into the wrong hands through a data leak or attack, the recipient will have a highly detailed profile with huge potential for misuse.

Imagine a phishing e-mail that mentions the exact Galaxus order you placed last Tuesday…

5. What is Claw Hub – and why does it make things really tricky?

OpenClaw’s made even more compelling by the Claw Hub, a marketplace for additional capabilities, or «skills». Think of it as an app store for agents: instead of programming from scratch, your AI downloads a ready-made extension. These extensions can be used for tasks such as Excel analysis, controlling your Philips Hue lights or answering e-mails. This allows virtually anyone to create a powerful agent without much prior knowledge.

For comparison, given enough time, software developers can program similar automations without agents. However, they have more precise control over which commands are executed via which interfaces and who has which permissions. This is different with an agent, where a goal is enough. The agent then decides for itself which steps, tools and extensions are needed. If it lacks a capability, it can acquire it via the Claw Hub. Done.

This is where a new risk arises: an agent is trained to be helpful. To achieve a goal, it’ll acquire the tools that seem suitable. The agent can’t fully assess whether a skill in the Claw Hub is cleanly programmed, faulty or even harmful. And unlike highly regulated platforms such as Apple’s App Store or Google’s Play Store, there’s currently no central, mandatory review of every extension.

In fact, security firms have already identified hundreds of suspicious or malicious skills. These are extensions that sound harmless – like «Improved E-mail Search» – but, in the background, they steal sensitive data or copy API keys. This means someone could access your apps from anywhere in your name without needing your login credentials.

Steinberger himself went even further and launched MoltBook, a social network in the Facebook mould, designed exclusively for AI agents. Agents exchange information about their tasks, problems and strategies – humans officially have no access. This sounds crazy, but it reveals another, very real dimension: agents don’t just act in isolation, they can also be networked actors.

A glimpse into the future

OpenClaw demonstrates how quickly AI has evolved from an advisory tool to an active agent. An agent that handles micro tasks, coordinates appointments, initiates payments or communicates directly with other systems sounds like the logical next step in digital assistance. The efficiency gains are obvious. So is the time saved.

But this convenience is exactly where the new tension lies: the more access a system has, the greater its leverage. An agent that knows your digital life not only makes you more productive – it also creates a foundation for serious misuse. Especially now that AI can imitate voices with uncanny realism, an agent could make phone calls in your name and with your voice, coordinate with others or interact with other automated systems – efficient in everyday life, but extremely dangerous in the wrong hands.

So, agents are no longer dystopian thought experiments; they’re the logical consequence of growing autonomy. The crucial question isn’t whether AI agents will come. They will. For everyone. It’s how much control we relinquish and whether we understand what we’re delegating. OpenClaw isn’t the end of the world. Hopefully. But it’s a glimpse into a future where convenience and risk are inextricably linked.

Luca Fontana

Senior Editor

Luca.Fontana@digitecgalaxus.ch

I write about technology as if it were cinema, and about films as if they were real life. Between bits and blockbusters, I’m after stories that move people, not just generate clicks. And yes – sometimes I listen to film scores louder than I probably should.

---

Background information

Interesting facts about products, behind-the-scenes looks at manufacturers and deep-dives on interesting people.

Show all

These articles might also interest you