E-Commerce Projects

Februar 2026

Agentic Workflows – my own openClaw

Workflow architecture 
Excited by AI and the initial reports about openClaw, I decided to take a look at it; I wanted to understand how it worked and what could be done with it. After the initial "wow" factor, however, came disappointment. In my view, the system is overly complex, involves a lot of overhead, and is incredibly powerful—perhaps dangerously so if you aren't careful. On top of that, it offers so many possibilities that it becomes overwhelming. I felt the same way I did back when the internet first became accessible to the average person: I’d launch the browser, stare at a blank page, and wonder, "What do I do now?"

It wasn't until search engines arrived that the World Wide Web truly opened up to me. I had a similar experience with openClaw: "What exactly am I supposed to do with this?" The few automation use cases that came to mind weren't things I wanted an AI to handle—and I certainly didn't want to send my data off to OpenAI, Claude, or anywhere else.

So, I examined how openClaw operates and built my own agent, "Bolle." Its architecture follows a similar pattern but is significantly simpler. Bolle’s core is written in Python and orchestrates various "skills" (others might call them "tools"—it makes no difference). Every request sent to Bolle is analyzed by an LLM. The LLM is instructed to identify the appropriate skill (or skills) for the request based on the available capabilities (which are part of the context). Each individual skill is also written in Python and uses input from the orchestrator (the core) to function. For instance, one skill might research the internet, while another could generate its own Python code if necessary. The skill's output is then fed back into an LLM call—potentially to trigger another skill that processes the previous result further. Finally, the results from the various skill executions are passed into a final LLM call to generate a coherent, well-formulated response. That’s it—nothing magical about it. After numerous attempts to run an LLM locally, I had to give up due to hardware limitations and am now using GPT-4.0-mini after all. It is relatively inexpensive and can actually provide structured responses (in JSON format) based on the provided context, making the output usable as input for the skills.

The bottom line: I really only use Bolle to remember things, research online, summarize content, or create lists. For anything beyond that, I don't trust the AI's accuracy enough to risk accidentally deleting or activating things I didn't intend to. And since Google Gemini or Claude can do everything Bolle currently does, for me, Bolle remains nothing more than a little toy.
However, it would truly come into its own if I gave it access to various APIs or system files. Then it could summarize long email threads, manage a shop system, build marketing campaigns, or generate analyses for me. If only I trusted the LLM enough...