Claws

I mentioned OpenClaw in my last post as something that finally achieved what I was trying to build with long-running autonomous agents. I've been running Clawdbot → BoltBot → OpenClaw since those early days. It's similar to my duyetbot-agent but much better executed - I still can't believe one person built something this complete, with proper design and documentation.

I ran the installation and let it configure itself - pretty good CLI. I deprecated my in-house duyetbot (the Claude Agent SDK version) and replaced it with the OpenClaw engine, giving it life through an SOUL.md that could be me (but then I realized that's better - it should be an assistant, not a virtual avatar).

I use OpenClaw pretty basically - mostly I let it:

• handle cronjobs, sending me daily messages, emails, and a rich website version with things to learn and read
• use tmux + coding skills to work with the installed claude
• maintain the homelab machines

Then I wanted to see how far it could go. I told @duyetbot to build a company.

It set up a Telegram group called Duet Company, appointed @duyetbot as tech lead, he onboarded a second team member running a GLM model. When they asked for a GitHub Org, I created one for them. I watched them build the company website, write docs, ship the first product. They brainstorm, they break things. I just watch and forward useful hackernews links when something seems relevant.

The fascinating part isn't that they can do this - we've known AI can write code for a while now. The fascinating part is that they're doing it autonomously, in the background. But so far things aren't as good as I expected - they still need human guidance. But this seems like the start of something bigger - AI capacity will grow later this year.

The claws family is interesting because it shows something happening in AI right now.