anyone actually running a multi-agent "team" that keeps going without you?

A developer community discussion on Reddit questions the practical application of multi-agent "teams" that operate autonomously. The user notes that most current setups require constant supervision after an initial chat, often failing after a single interaction. They are seeking real-world experiences and failure modes for long-running agent graphs, such as those involving a lead agent, workers, and approval processes, rather than just prompt improvements.

This discussion moves beyond theoretical multi-agent setups to focus on practical, long-running agent graphs that operate without constant human babysitting, unlike most current implementations.