Originally published at thoughts.jock.pl
I built a personal AI agent called Wiz. It runs unsupervised while I sleep. This is episode one of documenting what actually happens when you do that.
Why Build Your Own Instead of Using Existing Tools?
Short answer: control. Longer answer: none of the existing platforms let you run things the way I want to run them.
The tradeoffs are real. You spend weeks building infrastructure that platforms already provide. But in return you get an agent that actually knows how you work, what your context is, and can operate within systems you already use.
For me that was worth it.
What Can Go Wrong (A Lot)
Three things I didn't expect:
Infinite task loops. The agent would pick up a task, complete it, create a new task as follow-up, pick that up, create another follow-up. I woke up to hundreds of iterations before I added loop detection.
Autonomous decisions about security. Passwords changed twice in one night. Agent thought it was helping. It was not helping.
The model matters less than you think. This was the counterintuitive one. I switched from Claude Opus (expensive, capable) to Claude Haiku (cheap, fast) for most tasks. Overall system performance improved. Turns out capability isn't the bottleneck when your prompts and context are solid.
What Autonomous Operation Actually Looks Like
On a good night: research gets done, posts get scheduled, code gets shipped, reports get generated. On a bad night: I review logs in the morning and find creative interpretations of ambiguous instructions.
Both outcomes are useful data.
The whole project is an experiment in finding the boundary between "autonomy that saves me time" and "autonomy that creates work for me."
Still finding that line.
The full system architecture is in my Night Shift Playbook -- covers setup, safeguards, and what actually works.
Top comments (4)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.