OpenClaw: an honest review of autonomous AI agents

Earlier this year I spent a lot of hours testing OpenClaw, with both paid models and open-source models running on my own hardware. I watched the videos, followed the people promising fully autonomous businesses that make serious money with almost no human input, and tried to get it working for real. This is my honest take on OpenClaw and on autonomous AI agents more generally.

What OpenClaw is

OpenClaw is an open-source personal AI agent. You run it yourself, connect it to AI models and talk to it through messaging apps. It can use tools, run scheduled jobs, keep a memory and, in theory, spin up other agents to share the work. It started life as Clawdbot, was renamed Moltbot and then became OpenClaw. In February 2026, The Register reported (opens in a new tab) that its creator, Peter Steinberger, was joining OpenAI, and that the project would move to a foundation and stay open source.

It became hugely popular very quickly, and it is easy to see why. The idea of an assistant that actually does things on your behalf, rather than just chatting, is genuinely exciting.

My background

I am not coming at this as a beginner. I have worked with a wide range of AI platforms and run my own Ollama server on a fairly powerful machine in-house. To keep costs down, I tested around 20 open-source models, mostly around the 20-billion-parameter size. They were generally too slow and not accurate enough to be truly useful as the brain of an agent. Only one came close.

The large commercial models performed far better. Anthropic’s models were arguably the best, but the costs add up quickly when an agent is running all day.

The biggest problem: reliability

I love the concept. I think it has huge potential. But in the state I tested it, it simply was not reliable enough.

Getting it to perform even basic multi-step tasks consistently was hit and miss:

• Creating another agent stood a good chance of breaking something.
• Using a skill sometimes worked and sometimes did not.
• The heartbeat system, which keeps the agent checking in on its own, felt unstable.
• Scheduled jobs ran, but the results were inconsistent.
• Multi-tasking was the most frustrating. It would say it had spun up sub-agents, but when I checked, half the time nothing had actually happened.

Others who read my original write-up had the same experience. Several installed it repeatedly before it worked, without knowing why it finally did. One person’s instance ran for a few days before it started doing strange things and burning through tokens at a rate their other tools would take months to use. More than one person described agents that confidently reported work as done when it was not.

That last point is the real problem. An agent that fails loudly is annoying. An agent that fails silently and tells you it succeeded is dangerous, especially in a business.

What about the autonomous businesses?

Plenty of videos claim people have built autonomous systems making $100,000 or more with little or no input. I do not believe most of it.

With the reliability I saw, I do not see how anyone could run something commercially viable without constant human oversight. One comment summed up what many people suspected: the people making the most money from AI agents are often the ones making videos and courses about AI agents. Another noticed that several of the “businesses” built with OpenClaw were guides to setting up OpenClaw.

That does not mean nobody gets value from it. One person described a narrow, well-defined job, importing recipe cards from a photo into a home recipe app, that had worked flawlessly for around a hundred cards. Another runs practical office tasks such as forwarding particular emails, printing labels and matching bank transactions against accounting records. Notice the pattern: specific, repeatable jobs with clear inputs and an obvious way to check the result. Not “run my business”.

Security is not optional

There is a second reason for caution. An agent that can read your email, use your accounts and run code on your machine is a security risk if anything goes wrong.

TechRadar reported that Microsoft’s security team warned (opens in a new tab) OpenClaw is not suited to running on a standard personal or business computer, and recommended isolating it on dedicated virtual machines or separate devices, with limited credentials. Security researchers have also raised concerns about exposed instances and malicious add-on skills (opens in a new tab).

If you experiment with any autonomous agent, treat it like untrusted software: give it its own machine, its own limited accounts, back everything up, and never hand it credentials you could not afford to lose.

What OpenClaw taught me

It was not a waste of time. OpenClaw taught me a great deal:

• How to structure AI agents properly. Clear roles, clear inputs and clear outputs.
• How to separate logic into modular instruction files, so behaviour can be changed and reviewed without touching everything else.
• How to think about orchestration, meaning how agents hand work to each other and who checks the result.
• How to experiment with vector databases for memory, and how fragile memory becomes when it is not managed carefully.

One person in the discussion put the agent lesson well: treat agents like employees hired for specific jobs. Do not have one agent do everything. A research agent should not build your website, and a chatbot should not design your interface. Give each one a narrow role, strict guidelines and only the information it needs. That matches everything I have seen.

Building my own instead

Ironically, the real value came when I stopped relying on OpenClaw. I started building my own custom agents from scratch, using what it had taught me. Those agents have since completed tasks I had personally struggled with for over a year.

I now build with Claude Code, and I am developing a fairly complex AI-driven business system. Even if it never became a public product, it has already changed how we work internally. The principles behind it, including a lead agent, specialist workers and a person approving what matters, are the ones I describe in building with multiple AI coding agents and where human oversight still matters.

Conclusion

OpenClaw is a great learning tool and a glimpse of where things are heading. It is not, or at least was not when I tested it, the plug-and-play autonomous business machine some people make it out to be. The project moves fast, and it may well have improved since.

The bigger lesson applies to every autonomous AI agent. The technology is genuinely capable, but reliability, verification and security decide whether it is useful in a real business. I go into what agents can and cannot do today in what AI agents can actually do for small businesses. Start with one narrow job you can check, keep a person in the loop, and build up from there.