Handoff details

[S4mpl3r]

Hi there. Congrats on the new launches, really cool stuff!

I have a couple of questions (too lazy to dig in the source code to figure it out, so please bear with me):

1. Are agents stateful or stateless by default? Do they preserve conversation history or should we handle this ourselves?
2. When delegating to the handoff agents, will the conversation history be forwarded to those agents as is or will it be modified? (e.g. will the handoff tool call be in the history that the target agent receives). Asking this because I have experience building a custom handoff feature with the gpt-3.5-turbo family of models (back in ancient times lol) and I remember passing the entire conversation history between agents caused the model to hallucinate new functions (by for example, calling functions of other agents, etc)
3. What's the recommended method of passing back to the triage agent when working in a multiagent workflow? Take the following scenario as an example:

# Start of turn
User prompt -> Set a reminder for me
Triage Agent -> handoff to Reminder Agent
Reminder Agent -> Sure, can you tell me the date and time of your reminder?
# Turn is complete

What's the right thing to do when the user submits the response here? Should I directly pass it to the Reminder Agent and continue the conversation there (in this case, how and when should I pass back to the Triage Agent), or start again from the Triage Agent? We can probably make it work either way, but wanted to see if the framework has a best practice in mind in such scenarios.

[rm-openai]

1. Stateless. But the result object includes the input+output, and you can get the combined history for the next turn by doing result.to_input_list()
2. Yes, the entire history is included by default. You can filter the inputs though, and we have a built in extension to remove everythign related to tools - for the exact use case you mentioned. https://openai.github.io/openai-agents-python/handoffs/#input-filters
3. In cases like this, both agents should be able to handoff to each other. So you'd start every turn with the last agent that produced output, and any agent can handoff to other agents if it doens't know what to do.

Feel free to reopen for followup!

[xin-ran-w]

3. In cases like this, both agents should be able to handoff to each other. So you'd start every turn with the last agent that produced output, and any agent can handoff to other agents if it doens't know what to do.

Hello, do you know if there is an example for this? I can't directly complete this by setting the handoffs parameter in Agent class.

[Manouchehri]

@rm-openai

In cases like this, both agents should be able to handoff to each other.

How do we actually do this? I can't seem to find an easy way to "go back" to the triage agent again to answer the question.

Related: #771

[rm-openai]

triage_agent.handoffs.append(other_agent)
other_agent.handoffs.append(triage_agent)

that code + appropirate instructions in the prompt, will allow each agent to handoff to the other

[Manouchehri]

Thanks! I found this as well, might be useful for anyone else reading this: https://github.com/openai/openai-agents-python/blob/main/examples/customer_service/main.py