I have tried to look into the future in my first week with OpenClaw.

For the last 7 days I have played around with OpenClaw, the agentic AI assistant that is now the fastest growing open source project ever. Like in a dialog with ChatGPT/Gemini etc. you communicate with it via a chat interface, but the difference is that A: this agent works for you 24/7, even if you are not online, B: it learns and improves over time, and C: this thing can reprogram and extend itself, it can build its own code to solve tasks on its own.

I have been developing software all my life, I started with BASIC on a VIC 20 in 1982. Based on this lifelong experience in the last week I felt like a spoiled kid visiting a new, unlimited candy store. I have seen the future, regardless of how shaky things still are.

Here is my journey…

First, let’s make a few things clear:

You will find little need for this if your daily life does not require any automated tasks that require some “intelligence”. Not everybody needs an assistant, be it in flesh-and-blood or virtual. But I found that as soon as there was this intelligent always-on assistant available, the number of such tasks exploded.
Using OpenClaw is intentionally being kept rather technically complicated by the authors because there are a million ways of shooting yourself into the foot if you do not know what you are doing. Especially regarding security and privacy! Though having a few decades of software development experience helps, it is not completely necessary!
This is a bumpy ride: I easily had 200 crashes and reboots of the system in one week, plus 4 complete re-installs/starting from scratch. Sometimes it took 30 minutes to regain a responsive system after a crash (but the amazing thing is: OpenClaw has mostly fixed itself!).

Installation

Trust me, you do not want to install this in your LAN and clearly not on your day-to-day computer. In theory this agent could wipe your computer and sabotage your network.

I decided to keep this far away from my network, in the cloud. I am using an OpenClaw instance that runs on myclaw.ai. It takes a few minutes to set up and the smallest instance for US$19/month was already enough for all my experiments.

After installation I gave it a name (“Oli Püller”), told Oli what I expect from him and then let Oli set up his working environment. This all happens in plain language. Here are the steps I took:

I selected Claude Opus 4.6 as my model (expensive, but good)
Install Telegram: After this step all communication with OpenClaw went through a Telegram chat (don’t use WhatsApp!)
Install GOG skill to access Google Mail, Docs, Drive. BUT I am using a dedicated, newly created Gmail account for this. I would not consider handing over my personal 15-years-of-email inbox. But I gave it read-only access to my calendar for now.
Install an API key for Brave and Gemini search
Github integration (I am doing 2 backups per day of all my settings to the repository)

For each of these steps I sent Oli my request via Telegram (e.g. “I want you to access my gmail account xx@gmail.com”) and then Oli guided me through the complete setup, only asking for assistance with credentials or API keys that he could not get himself.

Setting up productive tasks

To give Oli the context of my daily work I instructed him to read through our website and blog. He did that and stored this information in his memory.

Now it was time to give Oli a few tasks. I asked him to set up three repeating tasks:

Daily Briefing (07:00 CET) — Spiegel headlines, weather forecasts, my calendar, topic reports (AI, CDR, EW, LongCOVID), Portugal news, Portuguese word of the day, today’s viral YouTube video
CDR Market Intelligence Brief (03:00 CET) — market data, deals, policy, tech updates with source links (tracks daily prices for ETS, CDR, etc. and plots a daily graph)
Weekly Portfolio Pulse (Monday 10:00 CET) — scan all 26+ CDI portfolio companies for news, funding, press, partnerships

For each task we had a multi-step dialogue where he showed me the latest results and I gave feedback on what I wanted differently. For example “make the email a beautifully designed HTML email” or “for each fact that you report add a deep link so I can dig into a topic” or “use our company logo”.

See the screenshots of the resulting “Weekly Portfolio Pulse” below. After just a few days I already love my daily personal newspaper that I can change and optimize just by saying what I want.

Just running Oli day-to-day costs about US$ 10 per day, as long as I don’t reprogram it substantially (then costs spike). I think the results are worth this money, I am already getting actual value out of it.

This is still early

I am barely 7 days into my agentic AI journey. Barely scratching the surface. Everything is experimental, it is more exploratory play than structured problem solving. By the way, that is how OpenClaw’s inventor, Peter Steinberger, also describes how this whole idea came to life in the first place.

I already burnt through 100 million tokens for almost US$500 in 7 days

In the last week I used a multiple of the AI power as I used in all my life before combined. The screenshot shows my Claude AI usage, measured and billed in “tokens”, of the most recent days. OpenClaw has used almost 100 million tokens. For the same amount of US$ 500 I could have had an OpenAI/Claude $20 standard account for 2 years.

I could also use cheaper models (and will likely do that later), but for this explorative phase I intentionally stayed with the most capable, most expensive model Opus 4.6.

What’s next?

Except for read-only access to my calendar my personal assistant Oli still has no further insights into my work and life. In theory assistant Oli could be a contact-keeper, sales-person, travel agency, chief of staff, executive assistant, etc. But before I let him more into my life the two of us need to go through a bonding and trust-building process first.

Let me end with two thoughts about the future of these agentic assistants: What is metaphorically happening is that I am using a 250 mm cannon to drill a hole for an m5 screw: Instead of writing code in Python, which the computer would easily understand, I am describing what I want and with immense technical effort and millions of tokens the assistant figures out how that will work and then write code to execute the task. My daily token usage went up by two or three orders of magnitude.

When most current AI users switch to this new paradigm the amount of necessary compute/inference will explode even more. I have my doubts that this will even be possible in the near future.

And, currently the AI companies sell us the tokens and AI subscriptions very much under price, they subsidize every request that you come up with by >90%. In normal reality my week would have cost US$ 5000 or more. I doubt that this will actually become a sustainable business model anytime soon. Even though eventually agentic AI might be able to replace traditional software, this will not happen overnight. The business model that I have spent most of my career with, Software-as-a-Service (SaaS), isn‘t dead anytime soon.

Nevertheless: As long as AI remains cheap, I am having fun with it! OpenClaw is just the beginning of this concept… Eventually such this could become some serious workhorse. Not tomorrow, but some day.