Agent calls a tool: transfer_to_<some_functionality> instead of just carrying out a conversation to capture data.

[apiltamang]

Question

Sometimes when I ask an agent to help me with a task, instead of helping with the task (which in this case is helping to fill out a form using a structured set of conversation states), the model attempts to call a tool like so:

2025-07-24 16:46:04 INFO     [Thread-2 (_setup_agent_session)] root: Agent started. Name: Game Order Agent
2025-07-24 16:46:04 INFO     [Thread-2 (_setup_agent_session)] root: event type received: raw_model_event
2025-07-24 16:46:04 INFO     [Thread-2 (_setup_agent_session)] root: event type received: history_updated
2025-07-24 16:46:05 INFO     [Thread-2 (_setup_agent_session)] root: event type received: raw_model_event
2025-07-24 16:46:05 INFO     [Thread-2 (_setup_agent_session)] root: event type received: history_added
2025-07-24 16:46:05 INFO     [Thread-2 (_setup_agent_session)] root: event type received: raw_model_event
2025-07-24 16:46:05 INFO     [Thread-2 (_setup_agent_session)] root: event type received: raw_model_event
2025-07-24 16:46:05 ERROR    [Thread-2 (_setup_agent_session)] root: Exception while processing event: Tool **transfer_to_game_order_agent** not found
2025-07-24 16:46:05 ERROR    [Thread-4 (_stream_to_openai)] root: Error sending audio. Exception: sent 1000 (OK); then received 1000 (OK)

The GameOrderAgent is declared like so: https://ctxt.io/2/AAD4IEHvFA

This gameOrderAgent is actually a handoff from a different customerServiceAgent who handled the initial caller greet like so:

game_order_agent = RealtimeAgent(
    name="Game Order Agent",
    instructions=open(f"{dir_path}/business_context_gameOrder.txt").read(),
    tools=[forward_call, capture_game_order_form, get_estimate],
    handoffs = []
)
customer_service_agent = RealtimeAgent(
    name="Customer Service Agent",
    instructions=open(f"{dir_path}/business_context_customer_service.txt").read(),
    tools=[forward_call, schedule_appointments],
    handoffs=[phone_order_agent, game_order_agent, item_return_agent]
)

Not sure why the model thinks it needs to call a tool. Initially the 'customer_service_agent' correctly handed off to 'game_order_agent' but now the gameOrderAgent calls a tool instead of just carrying along with a conversation.

Having dabbled some with the openai-agents-js, I know back there it mentioned that the handoffs or function calls are carried out by calling an internal transfer_to_ method. Thus I think this error maybe something systematic.

However, the model doesn't always end up like this. Sometimes it legit carries the conversation gathering important facts along the way.

Would appreciate any insight.

[seratch]

Hi, thanks for asking the question. I know this is a general comment on prompting for reliable agent behaviors, but details of metadata like name, description, and so on matter for making the agent's handing off even more predictable and reliable. So, simply passing sub agents do not consistently work, you may want to try handoff utility to set name and description, which are more suitable for the handoff scenarios, rather than relying on the sub agents' metadata.

[apiltamang]

Thanks for the insight. That is definitely very useful for doing handoffs more reliably.

However, the issue above is to try understand why the model decided to call a tool "transfer_to_game_order_agent". At this point, I am not even sure what transfer_to_game_order_agent exactly stipulates. Was the model trying to transfer to an agent by the name 'game_order_agent' ? That's weird because the current model IS the game order agent.

[seratch]

Thanks for clarifying. It sounds like either something that can be improved on your instructions and/or agent metadata or this SDK's bug, but with given information so far, I cannot point what's the cause of the unexpected behavior.