My “dumb” version of ClawdBot finished coding late Friday night, and I’m pretty sure it did a better job vibe coding than I could — though I’m far less confident that would be true if I were at the helm.
For anyone who hasn’t run into it yet, Clawdbot is an open-source agent: “a free and open-source autonomous artificial intelligence agent that can execute tasks via large language models (LLMs), using messaging platforms as its main user interface.” It runs on your own machine, borrows its brain from whichever model you point it at — Claude, DeepSeek or one of OpenAI’s — and you talk to it from a chat app, the way you’d talk to a person. It has only carried the name Clawdbot since January 2, having been Warelay in November and Clawdis in December. Three weeks old under its current name, and already finishing a coding job late on a Friday night.
The unattended part is the interesting part. With a human at the helm, the work bends toward what the human already knows. Which of the two produced the better code this time isn’t flattering, and the winking caveat is real — with me driving, the comparison might well come out differently.
Which raises the question: how much further along is this over babyAGI? lol
Some distance, and less than you’d think. BabyAGI was Yohei Nakajima’s little script from the spring of 2023 — born, in his telling, from wondering whether one could build an “AI founder” that could run a company autonomously. Its loop was one agent executing a task, another creating follow-up tasks from the result, and a third re-prioritizing the list. The project’s own README now describes that original version as having “introduced task planning as a method for developing autonomous agents.” That’s the lineage. Clawdbot has the same pulse of planning and acting and looking at what happened — with a far better brain behind it, real hands on a real computer, and a chat window where the terminal used to be.
The plan was to set this up on a MacStadium Mac mini — MacStadium hosts Macs for companies, and pitches the ability to “run AI workloads on Mac” — and, if it worked out, to make it a company resource. That plan hit a wall. Don’t try MacStadium for this: insufficient access. I’m repurposing an older machine to test now, and will report back.
