Feedback Makes the Conversation Useful

AutoGen’s feedback mechanisms answer a practical question: what should a system do when the current answer, context, or action is not enough? This chapter focuses on the point where a conversational workflow changes course—through a person’s judgment, a new retrieval result, an execution observation, or targeted grounding. The reader will be able to distinguish feedback that changes control from information that merely accumulates in a transcript, and to decide when an additional capability closes a real gap.

Human input is a control path, not a failure

It is tempting to describe an autonomous agent as one that does not need people. AutoGen takes a more useful view. Human participation is one of the capability sources available to an agent, and the framework lets developers configure when that source is requested. A human-backed agent can ask for input at selected rounds, allow the person to skip a turn, or operate with human participation disabled. Autonomy is therefore not a permanent label attached to the system. It is an operating mode chosen for a particular task and level of confidence.

That distinction matters because not all missing information is of the same kind. A tool can supply a calculation or an observation from an environment. A retriever can supply a passage. Neither necessarily knows whether the result answers the user’s real question or whether a proposed action is acceptable. A person can resolve ambiguity, provide domain knowledge, or notice that a technically coherent answer is aimed at the wrong goal. The person’s response is valuable not because it is magically error-free, but because it supplies a form of judgment that the other capabilities do not reliably provide.

The math problem-solving examples in the AutoGen paper make this path concrete. A student can work with an assistant agent in an autonomous setting, or the user proxy can be configured to request human guidance during the solution. In the more elaborate arrangement, the student’s assistant can invoke an expert when its own solution is unsatisfactory. The expert responds through a separate conversation supported by an expert assistant, and the resulting message returns to the student’s assistant, which resumes the original problem. The important operation is not “add an expert.” It is the controlled transfer of a question and the expert’s answer back into the conversation that owns the task.

This is selective escalation rather than a handoff of all responsibility. The student’s assistant remains responsible for continuing the main interaction. The expert is consulted for a defined uncertainty. The proxy provides a boundary through which a person can enter, and the returned message becomes context for a later computation. If the expert is asked too early, the system becomes needlessly dependent on human time. If the expert is never asked, the assistant may continue confidently through a mistake it cannot detect. The useful trigger is a condition tied to the task: an unsatisfactory solution, an unresolved ambiguity, or a decision whose consequences justify review.