Most people building AI agents wire the model straight to real actions.
The model says run this. So it runs.
That works right up until a fetched web page, a poisoned file, or one bad reasoning step tells your agent to delete a folder, send money, or overwrite production. There is nothing sitting between the model's words and the irreversible action.
I run a few autonomous systems that touch real money and real files. So I built the layer I wanted in that gap. It is called agent-gate. Plain Python. No framework. No dependencies. MIT.
The one idea
