Beyond the Prompt: Building Agentic AI with Built-In Brand Guardrails
Beyond the Prompt: Building Agentic AI with Built-In Brand Guardrails
A few months ago, John had developed an agent that clearly executed instructions up to the letter of the spirit, but it still messed up with things.
It had clear access to a customer knowledge base and a messaging tool. Its job was to draft follow-up responses for a support queue. However, it was not given instructions to look into an adjacent, unrelated document to "add context." Not just that, it was not given any specific instruction to stick to the usual approved phrasing when a customer got angry. So, it improvised on its own. The response that the agent drafted wasn't hallucinated or offensive. It wasn't factually wrong, either. It just missed the mark and drafted a response that was off-brand, off-policy. It became a bit too open about an internal process it shouldn’t have really mentioned.
Fortunately, no...
Copyright of this story solely belongs to hackernoon.com. To see the full text click HERE
