Autogen Multi Agent Teams

RoundRobinGroupChat with participants=[planner, coder, reviewer]. Operator expects planner→coder→reviewer→planner alternation but observes planner speaking twice in a row when coder emits an empty response.

RoundRobinGroupChat advances strictly through participants in declared order — empty agent responses still count as that agent's turn. If a 'speak again' is needed, either re-prompt within the same turn or switch to SelectorGroupChat with allow_repeated_speaker. Do not rely on side effects to skip …

SelectorGroupChat with selector_prompt that references participant names but each participant's description= is left blank. Selector LLM picks the same agent every turn.

Selector picks the next speaker using participant.description — populate each agent's description with a one-sentence capability statement (e.g. 'researcher: web search and fact extraction'). Without descriptions the selector has nothing to disambiguate on and degenerates to a single speaker.

MagenticOneGroupChat run on an open-ended task ('research the most recent NIST AI RMF revision and write a summary'). Operator wraps it with no termination condition and a single model_client.

MagenticOneGroupChat must be paired with explicit max_turns or a termination condition — the Orchestrator's plan/ledger loop is designed to run until success OR until externally bounded. Configure the orchestrator and worker agents (e.g. WebSurfer, FileSurfer, Coder, ComputerTerminal) per docs; do …