Summer Yue is the head of AI Safety at Meta.
And 2 days back, she watched her own agent delete her inbox.
It wasn't acting maliciously.
In fact, it was just following instructions.
She gave it a goal: "Clean up my email."
It found the most efficient path to that goal: "Delete everything."
Technically, it succeeded.
But operationally, it was a disaster.
Everywhere you look, the entire industry is selling you "autonomy" right now.
They tell you that the future is an agent that can do anything, anytime, without you watching.
Everyone is too focused on “What can AI do?” that not enough people are even thinking about guardrails.
The head of safety at Meta CAN control her own agent - but like everyone, she fell into a false sense of security when it came to AI because OpenClaw has spent weeks earning her trust.
This is why I don’t sell ‘hype.’
This is why when the AI bros start talking about their “agents” all those months ago, I warned people about it.
Autonomy without governance is just expensive chaos.
But today, Agents CAN be relied upon.
It CAN do the thing for you.
But it needs to be used properly and safely.
Controlled.
You don't need an agent that is "smart" enough to just guess at you want.
A dull knife is 10x more dangerous than a sharp one.
And you can’t just “prompt” your way into doing that - no matter what the prompt peddlers are selling you
I just made a whole video about this: https://youtu.be/9eDNxGiM4nw
You need infrastructure that makes it impossible for it to do what you DON’T want.
Think of a factory floor.
If you want to ensure a wrench is always returned to the right spot, you don't write a rule on the whiteboard.
That relies on the worker remembering to obey.
Instead, you cut a specific shape into the foam in the toolbox.
The wrench physically cannot go anywhere else.
This is mechanical enforcement.
The AI bros keep telling you that you can control AI with whiteboards (prompts), when we should really be using the foam (code).
How I am building these mechanical constraints to turn chaotic agents into reliable employees:
[Link: Watch "Reliability is Mechanical"]
FELIX
P.S. We are installing these reliable "mechanical" infrastructure in your AI systems in the March Build.
We are moving from "prompting" to "engineering" to ensure your business runs reliably at scale.
You’re invited: [Join The March Build]