How to Add Guardrails to a Local LLM Agent

Guardrails provide strong runtime control for AI agents, offering checkpoints to validate and modify inputs or outputs. This guide demonstrates how to implement a tool-output guardrail using the OpenAI Agents SDK with a locally-run Gemma 4 model via Ollama. A practical example involves building a customer support agent that could potentially leak sensitive data like fraud risk scores from a search tool. By implementing a guardrail, the sensitive information is redacted from the tool's output before it reaches the agent, ensuring the final response to the user is appropriate and secure.