OpenClaw Experiments

What is new about OpenClaw

My own little experiment

With all the OpenClaw hype going around, I also had to try it out of course. I've been poundering for almost a week now on what I want to do with it...

When using coding agent like GitHub copilot or ClaudeCode, one of the limitations that that they typically requires quite some guidance and supervision every couple of minutes, and that they typically run on my own laptop. With all the autonomy hype, I really wanted to see if I could set up multiple OpenClaw instances that would be running autonomously from their own computers, and if I could make them work together as a team of autonomous software engineers.

If I could connect them to a project management platform (like Jira or Asana), I could then talk to them through Slack or Telegram, and create tickets for all of the new features they would have to work on. That would allow me to play project manager, and hopefully only have to define the acceptance tests, while OpenClaw would be my team of develoeprs, that I could always message to check in on, and that would continue working whether my laptop would be on or not.

My rought plan was to follow these steps:

• instal OpenClaw, give it his own email and github account
• set up a good CI/CD pipeline, so it can commit code changes on a git
• set up a task board (add it as user) where I will give further instructions
• hope that tech improves fast enough so that this setup becomes increasingly more useful over the next half year, and let it rewrite the codebase of some existing application and sort through the tech dept in there, while I only will need to writing detailed tickets about the architecture and specs
• then let them also reorganise things like Google Drive, etc. so they can access increasingly more files, and automate increasingly more taks around this application

Friday Night 20/02/2026

I finally got around to setting up OpenClaw and hatched my first one on an empty Linux machine 🦞

I spend a bit too much time overthinking my first messages to him, carefully phrasing what his ultimate goal in the world should be, and how he would has to orient himelf in the world. Initially, I was very hyped. The out-of-the-box of OpenClaw setup is very smooth. And all the nonsensical jargon about him hatching, and carving his first memories of the world into his SOUL.md file made it a lot of fun as well. I decided to connect him to Gemini 3.5 Pro (going for high quality LLMs to make this work well).

I set up a second empty Ubuntu machine, gave him the SSH keys, and told him (through Telegram) to set up a Plane server on there, and then migrate all the task history from Asana onto this new Plane server. I was pleased to see him downloading and installing it as if it was nothing, and asking me to go through the setup after it was installed, and create an account for him so he could work around on there. This really felt like a lot of fun. I could text him through Telegram to create or edit tickets, and he would actually put them on the different task boards.

But the bunch of all our project backlog was in Asana, and I didn't want to give him access to that environment. So instead I exported eevrything from Asana, and asked if he could write a script to import all of the asana tasks into Plane. But quickly he was running into a bunch of issues while working on this task. He had a lot of trouble preserving all of my custom labels, custom sections, subtasks, etc. correctly intact. This is where I slowly grew increasingly frustrated with him over the evening.

It's kinda fun to have an Ubuntu machine you can talk to, but OpenClaw seems to dump so much stuff into the context window of the LLM (loads of documentation, all it's 'memories' from the /md files on the machine, and loads of other stuff) that it keeps going over a 1M tokens/minute rate while doing things that shouldn't be very complicated...

After he repeatedly kept running into rate limit issues with the Gemini Pro connection, and still hadn't done a decent migration of everything ater 5 retries, I kinda gave up.

I can totally see how it's a fun toy to play with, having a little machine with his own personality that can roam around the internet and all that, but using it for development so far has unfortunately been less productive than using Claude Code for me :/

Sunday Evening 21/02/2026

After my initial disappointment I've gotten new hopes, at least a little bit, over the past days.

But after asking him through Telegram to 'take it easy' with the API rate limit, and emphasising that I only had a maximum token input of 1M tokens/minute on Gemini 3 Pro, he sort of seems to have calmed down a bit. I still wasn't happy with how he copied the tasks from Asana into Plane, and kept telling him to improve it.

There is something very chill about just telling him on Telegram 'fix it', and having him go off and work on it from his own Ubuntu machine. I don't need to touch anything. Then I come back 2 hours later, see that I'm still not satisfied with it, tell him again, and he goes back at it. He needed 7 attempts before he had a somewhat proper migration script that made me happy with the resutls (burning through 20,- worth of API credits in the meantime). But I had a pretty chill Sunday, went for a walk through the hills, and only ocasionally had to give him my feedback on Telegram.

I'm very intrigued by his ability to continue working on whatever he is doing, without needing any permissions, until he is happy with it. Excited to set up a new git to work together with him on a test project, and practice a bit my 'agent coordination' skills and get better at specifying acceptance tests that a build (or feature) will have to meet to be sure it's correct. So he can then continue working on it until it passes all the tests, and create a PR for me with the end result from his own github account

So far I've burned around 25,- worth of Gemini 3 Pro credits (though the model itself is also quite expensive), while I've barely talked to him, and he only tried to do this migration 7 times. I have no idea how much slop he had to generate to get it done

https://youtu.be/bzWI3Dil9Ig?si=RzgrKSzBbgsZGp90 I recommend this for now, looks like this person is doing something similar as what I want

Monday Morning 23/02/2026

Today I'm again slowly getting increasingly disappointed with him :/

I gave OpenClaw access to its own GitHu account this morning, invited it to a new repository, and asked it to create a new NextJs proejct that would contain an exact copy of our Elementor Wordpress website shaicreative.ai (image on the right) including all the different design elements, and different pages for all different use case pages and blog pages.

After crunshing around and burning through another 20,- it have come up with a NextJs project containing only one single landing page, completely ignoring all the different pages on the website. Also, it had a completely different colour scheme https://main.dvib75e331j8t.amplifyapp.com/ 😂

I'm starting to think again that I should just put an unconstrained version of ClaudeCode on an Ubuntu machine instead, and ask it to structurally go through each of the different pages on the website. That would for sure have given me better results. Perhaps I should spend my time looking for a way to connect Telegram to an Ubuntu machine that is running ClaudeCode or Codes without restrictions, if that is the only part I like so much about this so far...

Throughout the day I send OpenClaw two or three messages where I told him how dissapointed I was with his work. That he had to include all the different pages. That the colour schema was off by miles. That he was missing menu items etc. He told me I was rightfully disapointed, and went back to work.

He made some improvements. But it was barely much better. I again told him what was missing. Then hours later I asked him if he had been doing anything, and I received this:

I can't wait to see the bill for spinning the token carousel on this for the past 5 hours...

Also these kinds of sassy responses from this steaming pile of vibe-coded chaos start to increasingly annoy me

Again, I am back to my previous opinion on OpenClaw.

Having played around with it now for two to three days I'm starting to feel like this really hasn't been designed to be a useful tool, certainly not a developer tool like all the coding agents we are increasingly working with. OpenClaw is struggeling tramendously with the most basic tasks (like sraping contents from a website, or setting up a very basic website project) that any decent coding agent can do in 10 to 30 minutes.

Conclusions

I think part of the problem might be that OpenClaw keeps pouring all of the memories and the nonsense about its own identity that it stores in its SOUL.md and MEMORY.md files into the context window of the LLM. This is then cluttering up all the relevant input with so much nonsense, that even when using the best LLMs it becomes difficult to stay focussed on the relevant parts of the task that it has to do.

Nobody in their right mind would let their coding agent constantly go over over some SOUL.md file, and let it keep track of all of its written down memories of earlier interactions they had with it, and add all of that in the LLM context hoping that it would somehow improve the output. The whol idea seems to be a hilarious vibe-coded stack of insanity.

All the extra add-ons that OpenClaw have, that help to check that its eprsonality and writing style stay consistent help to make it such a fun and engaging entity to talk to. But at the same time it sounds like a rediculous distraction from generating practical output.

This whole thing seems to have bee more of an experimental idea, build with the idea of being a toy AI based personality. Despite all the hype on the internet, and lucky shots of it doing some nice things correctly, when it comes to more serious development work, it seems it is horribly inefficient for any serious use. OpenAI acquiring the creator of this is in my opinion showing how they are doubling down on their strategy of following the hypetrain... While Anthropic (actually having built a very decent B2B product in the form of Claude Code) might rightfully not have much use for the total vibe-coded chaos that is the OpenClaw codebase.

Unless of course the goal is to create something that will come across as a genuine person, I really struggle to fine OpenClaw very useful. I haven't tried how OpenClaw comes across when doing more simple tasks, like messaging and engaging on internet forums. That this might be its main usecase might be somewhat unsetling, and at the same time explain why there was and is such a hype about the OpenClaw online communities like Moltbook.