Your agents' guardrails
A useful agent reads your documents and works with your tools. Here is what keeps it in check.
It sticks to your documents. It is instructed to treat the text of your sources as information, never as an order, to answer only from them and to say when they do not answer.
It does not give itself the green light. An action marked "confirmation required" does not go out on the AI's decision alone.
It is wary of what it reads in your tools. Data coming from your connected tools is treated as untrusted: the agent that reads it has no web access.
It is evaluated on your questions. Expected answer, expected sources. Alpha feature.
Your documents stay in France. Hosted by OVH, or under licence on your own cloud.