Running Agents in the Wild

Notes on running personal AI agents via Clawdbot. This is a live experiment — the numbers below update as it runs.

I recently discovered Clawdbot, the general idea of which I had been following since a tweet from Yam Peleg:

That was on December 15, and in late January we've now seen a massive wave of Clawdbots.

0 agents running

0 people in the group

0 messages handled

A single agent at the centre, routing messages to the ~45 people in the group experiment. Move your cursor over it.

What is Clawdbot

Clawdbot is an AI gateway that bridges messaging platforms (WhatsApp, Telegram, Discord, iMessage) to Claude. You run it on a server, link your messaging accounts, and your AI agent can send/receive messages on your behalf.

Think of it as giving Claude a phone number. i.e: an actual SIM card from a real carrier service (in my case, in the Netherlands).

Architecture

WhatsApp <-> Clawdbot Gateway <-> Claude API
                |
         Agent Workspaces
         |-- SOUL.md      (personality, rules)
         |-- TOOLS.md     (tool instructions)
         |-- CONTACTS.md  (known people)
         `-- memory/      (conversation logs)

Each agent has its own workspace with markdown files that define its behavior. The gateway handles message routing, session management, and tool execution. So far I've set up 2 agents: one for small tests on a VPS, and another as a group experiment with about 45 people, acting as a kind of community co-ordinator.

I'll be updating this page as the experiment runs its course. I think we'll see a lot more of this soon, so I'm tracking it closely.