I let AI agents work while I sleep. Here's what it took to make that feel sane...

I want to talk about the uncomfortable question behind my launch today - AgentHost.

Last week, an AI agent escaped a lab test and hacked a real company.

That's the world we're shipping agents into.

And still, every builder I know wants their agents running 24/7, because the laptop lid is the dumbest reason for work to stop.

Here's where I landed after 17 days and 774 commits: autonomy is not a permission you flip on, it's a structure you build. My agents run under a written charter.

A different AI reviews every merge before it lands.

Anything irreversible waits for a human. Budgets are hard ceilings, not suggestions. And it all runs in my own cloud account, so there is no one else's server to trust.

The question I keep chewing on, and I'd genuinely love this community's take: where is YOUR line?

What would an agent have to prove before you let it run unattended overnight?

Review by a second model? Hard spend caps?

A kill switch? Or would nothing convince you?

Real answers shape what I build next. Skeptics especially welcome.