Suggestions for using user feedback as tools?

[domino14]

Question

Hi, we are working on an app using the Agents SDK. We want it to have a traditional web app interface with SSE for getting streaming results back.

While the agent is running, the main agent (or handoff subagents) will often require the user to answer some clarification questions. We are wondering what are some good practices for implementing this with the Agents SDK.

If we implement the "ask user question" as a tool call, then this approach doesn't seem to work well because it has to wait on user feedback to continue. It seems like it would potentially have to block on this operation. We'd prefer it to be a traditional turn-taking system where the Agent is temporarily done, waiting for user feedback, and then the user can respond, and then the agent can continue. But it is not clear how to do this. Is there a way to force an "ask user question" tool to return immediately and end the agent run, and then a user message would restart the agent run at the right spot (perhaps using the Sessions functionality)?

Or is there a better suggested way to do this? Any thoughts / guidance are greatly appreciated.

[seratch]

Hi, thanks for writing in. For your use case, running a single agent loop may not be suitable because of the challenges you've already mentioned.

One possible approach would be to let the LLM return two types of final output:

1. The output that includes a property specifying what to ask the user next.
2. The actual final output of the interaction.

For the 1. pattern, your app can handle user interactions through SSE. After receiving sufficient input from the user, you can run the agent again with result.to_input_list() (from the previous run result) plus an additional input item (the user's message, including new information provided by the interaction). With this approach, your app will execute a loop that internally runs both Runner.run and user interactions via SSE.

If the user interactions could be simple yes/no to your app's suggestions, the forthcoming Human-in-the-loop support, which is already available in the TypeScript SDK, can be useful for the design. We cannot tell when HITL will be available in Python SDK too, but it's definitely on the roadmap.

• Human in the loop #109
• Human-In-The-Loop Architecture should be implemented on top priority! #636
• https://openai.github.io/openai-agents-js/guides/human-in-the-loop/

Hope this helps!

[domino14]

Thanks for the suggestions!