Using OpenClaw to Build a Better OpenClaw

My first impression of OpenClaw was that of being completely blown away and obsessed with using it. I also noticed a lot of frustrations right out of the gate.

• Starting with Opus, it became expensive fast, which wouldn’t be a problem except Opus seems to be the only functioning model that gets close to what I’m looking for in a tool like this.
• Sonnet does a decent job but it’s not that much less expensive.
• GPT-5.2 I couldn’t get working because if reasoning was turned on then I got a 400 error, apparently a known issue.
• Grok did okay for basic tasks but for anything more complex it failed utterly.
• I first tried switching models myself at different times and that became unsustainable.
• I asked OpenClaw if it had any functionality for automatic routing of models based on the task. I don’t remember what it told me, I can’t rely on what it tells me, but I know that it didn’t work.

I started thinking of a reliable system adapted to me that knows how I work, is customizable, and handles tasks like a company would. Instead of Opus — let’s say the CEO — always handling every request, it would instead get routed to the right agent with the right amount of capabilities at the right time, behind the scenes.

More to come later. Hope it doesn’t become obsolete in two weeks.