AI guardrails are often presented as a simple configuration problem: enable prompt-attack detection, block harmful content... and your agent is protected. In a real agent however a managed guardrails are only the beginning.
This talk follows the development of a Agentic AI project [https://github.com/msalanci/logs_talk_to_me/tree/v3], deployed in AWS Bedrock AgentCore runtime - a conversational agent investigating AWS logs using SQL query and AWS API calls.
Through specific failures we explore the point where managed guardrails stop being enough and where small, deterministic pieces of custom code become necessary, why system prompt cannot reliably prevent architecture leakage and how hooks can enforce rules before a tool is called or an answer is returned.
It is a practical guide for builders who want to understand what managed guardrails do well, where custom validation is unavoidable, and how to combine both without building an overcomplicated security system.

Target audience are agentic AI builders or anyone interested in guardrails for LLMs
Preferred talk is 30 minutes, but I can prolong it to 45 as well