Building a continuous, durable personal agent

In January 2026, I caught wind of a vibe on Twitter. Clawdbot was something that people were talking about. Moisturize, they said. Exfoliate!

There was humour, AI, and a hint of something truly powerful.

First taste of lobster

Like many, I said YOLO! Fully aware — and yet in blissful denial — of what it means to run an unsandboxed agent on my personal machine. I called it Chonk and gave it a backstory about an obscenely large right claw (but don’t mention the left one). I set up the Realm of Chonk on Discord and was immediately blown away.

The interaction pattern of a continuous agent was fundamentally different from anything I had experienced before.

It could just build stuff and do stuff. Really, anything I could do from the command line was within its grasp. I started pushing it. Burning the midnight oil and tokens alike.

I wanted to know how it worked. I wanted to build my own. Not by studying the code, but by studying the interactions and figuring it out myself.

Of course I had Chonk doing this for me. It naturally had access to its own codebase and an infinite appetite for new challenges. And challenges I gave it. “Build an AI agent like ClawdBot but different in these ways…”

And it worked! I actually had my own AI agent that connected to Discord and with which I could interact. I needed to tell everyone about how incredible this was. I don’t have a huge audience, but if you were in my sphere of influence around January, you know.

But as these things go, I soon realized that I had over-extrapolated.

Cracks in the seams

Playing with the ClawdBot of that era (then MoltBot for a tragic moment and then OpenClaw) exposed some subtle cracks. Cracks that bugged me enough to think that I could fix them if only I built my own version. With Chonk just a message away, I was sure that I was just a few prompts away from what I wanted…

Narrator: No amount of mindless pulls at that slot machine’s lever were going to make a better claw pop out. This was a job for real engineering.

The things that were bugging me were the following. These frustrations directly influenced my vision in building TorkBot.

1. Continuity - My conversations in threads, the channel and DMs were basically with different agents.
2. Responsiveness - If the agent was working on a big task, it was effectively blocked until it yielded its turn.
3. Reactivity - If I wanted to steer it mid-task (“STOP DELETING MY DISK!!!”), I couldn’t.
4. Resumability - If I closed my laptop, if Clawd crashed, or if I had a hiccup w/ my inference provider, all the intermediate work was lost. I’d have to say, “pick up where you left off.”

I bet many, if not all, of these are now solved problems with the incredible team and systems that Peter has put in place around OpenClaw. But for me, I was completely and utterly Nerd Sniped.

The TorkBots you’ll never see

To build TorkBot as it is today, I followed quite a few dead ends. I had a few generations of TorkBots that were cool enough to keep me hooked but that just didn’t work in some critical way or another. You’ll never see them because they weren’t worth seeing. But I learned from each iteration and each dead end. Perhaps one day, I’ll do some git archaeology and try to recover the issues and the designs, but not today.

TorkBot today has achieved, to a small degree, many of the design goals I set out to tackle. But there’s always more. There is so much that an autonomous system that acts with semantic intent could be — and will be.

Today, TorkBot is my personal assistant and helps me in big or small ways. Tomorrow, though, Tork could be helping millions. Who knows?

Stick around to find out!