What I did
I set up an AI "director" to run all my projects. The idea was simple: I talk to one AI, and it handles the rest. Less time, more peace of mind.
To make it fit, I changed most of my working setup at once. It happened fast.
A few days later I was putting out its fires.
It spent most of its time managing itself. Its own improvement plan went through 7 review rounds, and the last one made no progress. Two jobs I actually needed got killed halfway because my laptop ran out of memory. It left processes running in the background, and I had 132 dead sessions to clean up.
I trusted it too much, too early. And since I'd changed everything in one go, there was nothing steady to fall back on.
Why a small test wouldn't have saved me
My first thought was "start it on small tasks." But it does small tasks perfectly. It broke under load - many jobs at once, running for days. A small test passes, you trust it, and the fires start the day you scale up.
The bugs weren't the problem. Every new system has bugs. The problem was that its bugs landed on the setup my business runs on.
What I'd do now
- Give it real work, in real amounts. That's how the bugs show up.
- Run it in its own space - a separate folder, account or test setup. When it breaks, nothing you rely on breaks with it.
- Step up the load. One project, then three, then all.
- Keep the old way running the whole time. Switch it off only after the new one holds up at full load.
Something to steal
Before you hand an AI agent anything new, answer these 3:
- If this breaks tonight, does anything I already rely on break with it?
- Am I testing it at the size I'll actually use it?
- What's my old way, and is it still running?
If the answer to #1 isn't "no", it's not ready.