everyone wants an AI agent to run their company.. here's how you build one that doesn't torch the place:
1. read access now, write access is earned - give it eyes on everything day one but the hands come later.
2. make it log everything - action, confidence, outcome, for 60 days before it auto-executes anything. same guardrails you'd put on a new hire.
3. every fact it acts on has to trace back to a source - no link, no action. a hallucinating agent in prod is scarier than any intern.
4. the failure isn't a weak model.. it's too much autonomy too early.
5. shrink the job till it can't fail. the fix is always tighter scope, not more capability.
..that's how AI actually runs a company.