Guardrails

Agents are useful because they act; they are risky for the same reason. A FAQ agent will happily answer questions it should refuse, and a helpful assistant will promise a deadline extension it cannot grant. Guardrails are the checks that run before the agent sees a request and before the user sees a response.

Two AISL workshops build the same safety flow from different angles, both around the DataTalks.Club Zoomcamp FAQ agent.

The OpenAI Agents SDK version

Guardrails for AI Agents with the OpenAI Agents SDK is the complete SDK walkthrough:

• Input guardrails block off-topic questions before the agent runs.
• Output guardrails catch policy problems — promising deadline extensions, writing homework for a student — before the response reaches the user.
• Tripwire handling turns a blocked run into a normal message.
• Multiple guardrails run in parallel; streaming changes when a block can still happen.

The framework-agnostic version

Build an Agent with Guardrails rebuilds the flow without leaning on SDK magic: the guardrail as a plain tool plus an asyncio runner, so the same pattern works with any framework. The async primer matters — guardrails only stay cheap if they run alongside the agent, not in front of it.

The January workshop closes the loop on cancellation: when a guardrail trips, the wasted agent run is cancelled mid-flight instead of finishing and being thrown away.

Where they sit

Guardrails are one layer of agent quality. What to build next, and how to know the guardrails and the answers are any good: Evaluation and Monitoring.

Where to go next

• The agent being guarded: Agents
• The FAQ agent these workshops guard, deployed: Agent Deployment