I tried openclaw a few weeks ago. It's pretty impressive at first and I see why there's so much hype around it... in the end it just a more versatile version of claude code.

I think it's importand to understand that "agentic ai" is still just an LLM predicting output just wrapped into an interface that provides context and tool calling in an "agentic loop".

It is pretty powerful, but it's still very prone to the usual LLM shenanigans, which I would describe as overconfidence. I found myself micromanaging my openclaw assistant a lot... but maybe that's also a trust issue on my side :P

and all the compute such an assistant needs is not to be underestimated... after a few hours of playing with openclaw I had burned several million tokens with anthropic!