OpenClaw is amazing, but after 2 weeks I'm back to manual – the reliability gap is real

Don't get me wrong, I love the idea. Installed it day one, played with skills, even wrote a few custom ones.

But on my actual messy desktop (20 Chrome tabs, 3 Office apps, VPN, weird folder structures), it fails ~40% of the time.
Not because the model is dumb – because the execution layer has no tolerance for real-world chaos.

• A popup appears -> agent freezes

• File name typo -> agent gives up

• App not in foreground -> click misses

I know I could write more robust skills, but that defeats the “voice and forget” promise.

Is anyone working on a harness layer specifically for desktop reliability? Something that retries, asks clarifying questions, and logs everything?
Feels like the model part is solved, the engineering part is not.