AI Agent Guardrails Why Your System Prompt Is Not a Barrier

Designing Agentic Systems | Episode 8 On the Mumbai to Pune ghat road, every sharp curve has two things: a sign that says drive slowly, and a steel barrier. Only one of them stops a truck. Most teams building AI agents say "we've added guardrails" when all they've done is write a few lines in the system prompt. That's a sign, not a barrier. This video shows where the real barriers go. In this video: Why an instruction in the system prompt is a sign, not a guardrail, and why text can always be argued with The one definition to carry: a guardrail runs outside the model, and it can say no What goes in: scope checks, masking card numbers and CVVs, and content classifiers (and why a classifier is a filter, not the last line) What gets done: structured tool schemas check the shape, the record check catches the wrong facts, and the gates check the permission What comes out: is the reply true, is it allowed, is it masked? And the three honest options when it isn't The whole picture: the model sits outside your code, and a policy service decides what's allowed Key takeaway: if the only guardrail your agent has is in the system prompt, it doesn't have guardrails. It has signs. Next episode: Episode 9, Retries and Failure Modes. Guardrails decide what an agent is allowed to do. Next time, what happens when something it's allowed to do goes wrong, and why you can never retry a payment the way you retry a web page. About this series, Designing Agentic Systems: Everything You Know About System Design Still Applies What an LLM Agent Actually Is Agent vs Workflow vs Chatbot Anatomy of an Agent Single Agent or Many Model Context Protocol Prompt Injection Guardrails (you are here) Retries and Failure Modes State and Memory Agentic AI in Banking Cost and Scale Ehson Golandaz is an enterprise architect with 23+ years in banking and financial services, and the author of two technology books. This series builds, episode by episode, towards a working agentic AI system for an enterprise. Not a demo. Something that survives production, audit and a regulator. #AIGuardrails #AgenticAI #AIagents #AIsecurity #EnterpriseArchitecture #BankingTechnology #LLM #SystemDesign #FinTech #ResponsibleAI