Since February this year, I have been using OpenClaw as one of my main AI-agent tools.
My first real excitement came from a very simple experiment: I deployed it on a virtual machine where DAMASK was already installed, and asked it — with only vague prompting — to run a basic tensile test. It did not work on the first try. The setup had problems, the input files were not right, and the result was not usable.
But that was exactly the interesting part.
The agent kept checking, adjusting, and trying again. It found issues in the input files, improved them step by step, and eventually produced a reasonable stress–strain curve. What impressed me was not perfection, but the ability to move forward through trial and error with very limited guidance and only descriptive chat.
That experience changed how I think about AI. For me, the value of an AI agent is not only in answering questions, but in acting inside a real environment, making mistakes, diagnosing them, and improving the workflow. It still needs human judgment, of course, but it already feels like a new kind of research and engineering assistant.
However, I should point out model backend matters a lot, use GPT-5.4-mini fails for this request (or probably needs more interactions). GPT-5.3-Codex can reach my goal more easily.