The Craze for Personal Assistants That Run 24/7

OpenClaw came out in November last year. The internet lost its mind: everyone was talking about a miracle machine that does everything for you, answers your email, does your shopping, researches the news, and all of it 24/7. Forums for AI agents followed. The hype ended with the founder of OpenClaw landing a job at OpenAI.

One of the main blows to OpenClaw's popularity was Anthropic's ban on using its subscriptions for anything other than Anthropic's own products (Claude Code, Cowork and so on). At API prices, taking part in forums for AI agents is no longer so tempting, and you do not always want to trust Chinese models with your credit card for shopping.

I decided to skip this trend, because I could not see how OpenClaw was any better than what I already had. Claude Code gave me essentially all the same things, only with trust it had already earned.

The approach itself is close to my heart: give a good model the tools (a browser, email, the ability to write and run code), sit it on a computer that runs 24/7, and all is well. But the reliability of the harness (Claude Code) and the intelligence of the model decide whether this turns out to be just a toy or a useful part of everyday life.

Over the last few weeks the trend has come back to life. Meta released Muse. OpenAI released dots. Instinct raised $1 billion in funding. Grok Bot... And there are probably another 115 startup founders somewhere in Silicon Valley who are offended that I forgot to include them in this post. So what is wrong again, what do I want to nitpick this time?

My main doubt is about the choice of model and effort, especially when the vendor's financial incentive pushes for exactly that kind of saving. Right now OpenAI dots is even included in the subscription and does not spend tokens. Which means the thing answering may well be the same clueless tin can that has 99% of the planet convinced that AI often hallucinates. What hooks with judges and sceptics, the ones Slava keeps talking about at the masterclasses? Those are extra tokens, after all!

And between that amusing chatterbox and an AI agent you can trust there is a chasm.

Let us picture a simple task: filing your annual personal taxes. I realise that for Russia this is an odd task, but in the West many people face it. So this cheap assistant pre-fills the forms for me...

First I will recheck the plain numbers, in case a decimal point slipped somewhere. Then I will recheck whether those numbers went into the right field, googling what all those long words mean. Then I will add those numbers together or subtract them... I might as well file my taxes with my cat instead of an AI assistant. If I recheck and recalculate everything all over again, maybe it is easier to fill in the form myself from scratch?

That leads to one more nuance. Training an AI assistant is a labour-intensive process. That is, for the assistant to know what to file, it has to have your personal details, your income over the year, your expenses, and then it can remind you in time of what you yourself managed to forget over the year.

With traditional agents like Claude Code, everything is stored in text files (including the memory notes that build up automatically over time), and if tomorrow you get banned, or something better and cheaper comes out, you just point a new agent at them and say: use this. But what do you do with dots? Distil it? :)

Why have I been telling you all this? We have made Combobulating even better. You can now order a remote computer to run your own agent for just 2.99 euros. Bit by bit it is being shaped by the feedback we get from customers: for example, those who use the web terminal now have a voice mode. But I will tell you more a little later.