Guardrails & Tracing

Agents fail silently without traces, and one bad output can leak data or violate policy. Guardrails + tracing are the minimum production safety net for OpenAI Agents SDK apps.

Guardrails are airport scanners; tracing is the black box flight recorder — one prevents bad things from passing, the other explains what happened when something goes wrong.

Scenario

Output guardrail blocks responses containing credit-card patterns; trace shows the Billing Agent called `lookup_invoice` twice before succeeding.

Solution

In Agent Frameworks, apply Guardrails & Tracing to this scenario: Output guardrail blocks responses containing credit-card patterns; trace shows the Billing Agent called `lookup_invoice` twice before succeeding. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Do this before moving to the next module — reading alone is not enough.

Open the Code Walkthrough below and run it locally. Change one parameter related to Guardrails & Tracing (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Highlighted lines show where Guardrails & Tracing happens in the code.

1from agents import Agent, Runner, trace, OutputGuardrail  # import dependencies2
3def no_pii(ctx, output):  # define a reusable function4    if any(c.isdigit() for c in output[-20:]):5        return GuardrailResult(tripwire_triggered=True)  # return the result6    return GuardrailResult(tripwire_triggered=False)  # return the result7
8agent = Agent(name="Support", output_guardrails=[OutputGuardrail(no_pii)])  # key line for Guardrails & Tracing9with trace("support"):10    Runner.run_sync(agent, "What's my SSN on file?")

• Treating Guardrails & Tracing as a black box without evaluation
• Ignoring cost and latency in production
• Skipping error handling for openai guardrails tracing

Quick recap — the most important points from this module.