Since assistants_agent might handle other sub-tasks and needs to output the results of those sub-tasks, I have to give it control of the conversation.
Maybe I can pass the output of assistants_agent to query_agent as well, and let query_agent make the final output ?

Sorry, I just figured out what handoff does, it can solve this problem

Although I solved this problem, I still want to know why they behave differently. When using 'query_agent.as_tool', the instructions in query_agent do not seem to take effect.

hi, I created a query_agent that can invoke tools. There is also an assistants_agent that has control over the conversation.

If I use handoffs = [query_agent] in the assistants_agent, it can invoke tools well and pass the correct parameters, but this causes the assistants_agent to lose control of the conversation.

I tried using tools = [query_agent.as_tool] in the assistants_agent, but the performance of invoking tools deteriorated, and the parameters passed were not always correct.

Is this normal? How can I solve it?

Hi @huangbhan, thanks for sharing this — I'm in the same situation.

Could you please explain the difference between using handoffs=[query_agent] vs tools=[query_agent.as_tool()]?
And what are the practical consequences of choosing one over the other (in terms of behavior, control flow, parameter passing, etc.)?

Thanks in advance!

there's great explanation here: https://cookbook.openai.com/examples/orchestrating_agents

The basic idea is handoff share the conversional context as the last agent, but tool calls don't.