A Python agent built with Microsoft Agent Framework has a refund_payment function tool.

How it works

In Microsoft Agent Framework (Python), you mark a function tool as needing human approval with the @tool decorator: approval_mode="always_require".

What happens next is the part people get wrong. The agent doesn't pause inside the run:

1. Instead of executing the function, the agent might respond with a request for approval. The run completes with a response that says what input is required from the user, instead of a final answer.
2. The response carries user_input_requests. The caller (your app) is responsible for getting the decision from the user.
3. You build the reply with to_function_approval_response on the user input request, and pass it back to the agent as part of a new agent run.

The samples loop on this: while the result still has user_input_requests, collect approvals, run again, and stop when none are left.

On never_require, the docs are explicit: samples use approval_mode="never_require" for brevity, and you should use "always_require" in production.

The other options, and when you'd use them

• The framework prompts the user itself: The caller is responsible for getting the user's input and passing it back.
• The same run pauses and continues: The run completes with the input request, and the approval goes back in a new agent run.
• never_require plus instructions: Instructions don't enforce anything. never_require is for sample brevity, and production should use always_require.

Exam tip

always_require -> run returns user_input_requests -> caller collects the decision -> to_function_approval_response in a new run.

References

• Using function tools with human in the loop approvals