Maintain State for Multi-Turn Slot Filling in a Multi-Agent Handoff Scenario

[tnduyh5]

Hi,

I’m building a multi-agent workflow with the OpenAI Agents SDK and am running into a routing/state issue during slot filling after a handoff.

Setup

• Main (Orchestrator) Agent
  Role: Classifies the user’s intent and delegates via handoff to the right specialist.

• Rental Car Agent (Specialist)
  Role: Handles rental car bookings.
  Tool: make_car_reservation(location: str, date: str)
  Behavior: If location or date is missing, the agent asks follow-ups until all required slots are filled, then calls its tool.

Problem (Current vs. Expected)

• Current (Incorrect)

  1. User: “Hi, I’d like to book a rental car.”
  2. Main Agent routes via handoff to rental_car_agent.
  3. Rental Car Agent: “Where would you like to pick up the car?”
  4. User: “I’ll be in San Francisco.”
  5. Bug: The Main Agent treats this as a fresh query and replies itself (breaking the flow), instead of letting the Rental Car Agent continue slot filling.

• Expected (Correct)
  After the initial handoff, subsequent user replies in this session should be sticky to the Rental Car Agent until it either:
  a) completes the slot filling and calls make_car_reservation(...), or
  b) explicitly hands back control to the Main Agent.
  In other words, routing should bypass general intent detection while a specialist-led slot-filling routine is in progress.

My Questions

1. Forced Routing During Slot Filling
   What is the recommended pattern to guarantee that user replies are routed back to the same specialist agent after a handoff—until that agent finishes or explicitly returns control?

Thanks!

Debug information

• Agents SDK version: 0.2.11
• Python version: 3.11

[ihower]

@tnduyh5 Since you didn’t include a reproducible code sample, my guess is that on the next turn you’re not passing the latest agent back into the runner.

Here’s a working example that shows the pattern:

from agents import Agent, Runner, function_tool
from agents.extensions.handoff_prompt import RECOMMENDED_PROMPT_PREFIX

reservation_agent = Agent(
    name="Reservation Agent",
    instructions=f"""{RECOMMENDED_PROMPT_PREFIX}
You are a professional reservation specialist handling reservation requests efficiently.
When customers ask about topics unrelated to reservations, transfer the conversation to TriageAgent""",
    model="gpt-4.1"
)

triage_agent = Agent(
    name="Triage Agent",
    instructions=f"""{RECOMMENDED_PROMPT_PREFIX}
You are the main triage specialist responsible for routing customer requests.
""",
    handoffs = [reservation_agent],
    model="gpt-4.1"
)

reservation_agent.handoffs.append(triage_agent)

async def main():

    messages = []
    current_agent = triage_agent  # start with orchestrator

    while True:
        user_input = input("User: ")
        
        if user_input == 'q':
          break

        messages.append({"role": "user", "content": user_input})

        result = await Runner.run(current_agent, input=messages)

        print(f"{result.last_agent.name}: {result.final_output}")

        # Update conversation state
        messages = result.to_input_list()

        # 🔑 Key point: after a possible handoff, always use the latest agent
        current_agent = result.last_agent

if __name__ == "__main__":
    import asyncio
    asyncio.run(main())