Implementing Guardrails for Agent Behavior Based on OpenAI Guidelines

[mayestik1]

Hello everyone,

I’d like to open a discussion regarding the implementation of guardrails for agents in this project, in line with the best practices suggested in OpenAI’s recently released guide:
A Practical Guide to Building Agents (PDF)

The document outlines key principles for ensuring agents act safely, predictably, and within intended constraints. I believe integrating such guardrails—whether through input/output validation, permission systems, monitoring, or role-based control—would enhance both the reliability and safety of agents developed in this ecosystem.

Would love to hear your thoughts on whether and how we might incorporate this into the project.

Looking forward to the discussion!

[ilvalerione]

I'm not sure how this can be implemented in practice, but I think we are underestimating the role of the system instructions here.