OpenAI Agents SDK

• Key points are in the visual diagram above.

Building agents on raw chat completions requires manual loop management, tool routing, and guardrails. The Agents SDK provides primitives — Agent, Runner, Handoff, Guardrail — that handle orchestration, letting developers focus on agent logic and tool design.

The Agents SDK is like a call center management system — it routes calls between departments (handoffs), enforces scripts (guardrails), and logs every interaction (tracing), while agents focus on solving problems.

Scenario

Customer service system: triage agent classifies query → handoffs to billing agent (has payment tools) or tech agent (has diagnostic tools). Input guardrail blocks PII. Output guardrail ensures professional tone. All steps traced in OpenAI dashboard.

Solution

In OpenAI Agents SDK, apply OpenAI Agents SDK to this scenario: Customer service system: triage agent classifies query → handoffs to billing agent (has payment tools) or tech agent (has diagnostic tools). 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 OpenAI Agents SDK (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 OpenAI Agents SDK happens in the code.

1from agents import Agent, Runner, handoff, GuardrailFunctionOutput  # import dependencies2from agents import input_guardrail, output_guardrail  # import dependencies3
4@input_guardrail5async def pii_guardrail(ctx, agent, input):6    if contains_pii(input):7        return GuardrailFunctionOutput(tripwire_triggered=True, output_info="PII detected")  # return the result8    return GuardrailFunctionOutput(tripwire_triggered=False)  # return the result9
10billing_agent = Agent(11    name="Billing Agent",12    instructions="Handle billing inquiries. Use payment tools.",13    tools=[check_balance, process_refund],14)15
16tech_agent = Agent(17    name="Tech Agent",18    instructions="Diagnose technical issues. Use diagnostic tools.",19    tools=[run_diagnostic, check_status],20)21
22triage_agent = Agent(23    name="Triage",24    instructions="Classify the query and hand off to the right specialist.",25    handoffs=[billing_agent, tech_agent],26    input_guardrails=[pii_guardrail],27)28
29result = await Runner.run(triage_agent, "I was charged twice for my subscription")

• Too many handoffs — latency and context loss
• No guardrails on production agents
• Not using tracing — can't debug multi-agent flows
• Monolithic agent instead of specialist handoffs
• Ignoring max_turns — runaway agent loops

Quick recap — the most important points from this module.