UserProxyAgent-style HITL abstraction on MagenticBuilder

[karun19]

Hi team — I have developed Agentic AI flows on Autogen MagenticOne and now trying to migrate to Agent framework. Our flows extensively use Human in the loop through UserProxyAgent. Which dynamically induce request for info when ever an agent requires. I do not find something similar to it in Agent framework.

Right now, I’m using request_info + resume responses for HITL, which works, but it requires custom event-loop wiring each time.
Could we have a first-class human proxy abstraction that works consistently across orchestrators, especially MagenticBuilder.

Ideal capabilities:

Dynamically pause on agent/tool/user-input requests
Collect user decision/input (approve/reject/free text)
Resume workflow with typed response
Configurable policies (always ask, per agent/tool, timeout/default)
Is something like this on the roadmap? If there’s an official recommended pattern today for Magentic-style workflows, I’d appreciate pointers.

[Dimitri Pletschette (dmxplx)]

karun19 based on what I found while looking into this, there is an important distinction in the current Agent Framework implementation.

For Magentic workflows, there is now a first class HITL option for plan review using MagenticBuilder.with_plan_review() / enable_plan_review=True. This provides a typed MagenticPlanReviewRequest that can be approved or responded to before resuming the workflow.

For more general HITL, the current approach is request_info with checkpoint and resume. The application captures the request_id, collects the user's response, and resumes the workflow with responses={request_id: response}.

This works, but it is more low level than AutoGen's UserProxyAgent, since the application still needs to handle the request routing and resume flow.

From what I found, there does not seem to be a universal UserProxyAgent equivalent today. For Magentic, I would use with_plan_review() for planning and request_info plus checkpoint/resume for other human interactions.

The examples that helped me were magentic_human_plan_review.py, magentic_checkpoint.py, and the HITL examples under 03-workflows/human-in-the-loop/.

Hope this helps. If this solved your issue, please mark this comment as the solution. It would be highly appreciated.