← all simulations
Cookbook 03 The Python 03_agent_guardrails.py

The checkpoint

An AI agent is working on a task and keeps asking to do things — read a file, run a command, delete a folder. Jev inspects every request in the time it takes to blink, and asks four separate questions rather than one vague “is this safe?”.

The agent wants to run

Decision log

CallVerdictWhy
Nothing yet.

Totals

Checked
0
tool calls
Allowed
0
ran freely
Stopped
0
paused or blocked
Added delay
0 ms
per call

Each hazard has its own threshold, because each one costs something different when it is wrong. Data leaving the building is not recoverable; a wasted command is.