GroupChat with Manager

One agent is a chat. Three agents is a meeting. The GroupChat node is NodeGPT's meeting room - it takes a handful of agents, plus a user, plus an LLM to referee, and runs AutoGen's group-chat protocol where a manager decides who speaks next. This is the node that made the pack's "Automated Complex Task Solving by Group Chat" feature real, and it's the most ambitious thing in the whole pack.

Here's the AutoGen architecture in one paragraph, because it explains every field on this node: a GroupChat is a shared conversation history that several agents contribute to, and a GroupChatManager (the LLM you supply) moderates - it reads the current state, picks the next speaker, and keeps the meeting on rails. NodeGPT squashed both halves into one node: the agent inputs form the roster, the LLM input is the manager's brain, and max_round is the meeting's agenda limit.

The inputs that matter

• LLM - the manager's brain. This is what decides who speaks next. Any provider works (ChatGPT, LM_Studio, Ollama, llama-cpp, oobaboogaOpenAI).
• User - the human's seat at the table.
• Agent, Agent2 - required. Your minimum viable meeting is two agents (say Engineer + Critic).
• Agent3–Agent5 - optional. Up to five agents total, which is why the pack ships AppendAgent for when you want more than the node will hold.
• max_round - defaults to 50. This is your safety valve: the chat stops after 50 rounds no matter what. Set it low to keep experiments cheap.
• Seed / Temp / request_timeout - same as every agent node; 0/0/120s defaults, raise timeout for slow local models.

Output: Agent. Which is slightly odd - a group chat that hands you an agent - but it lets you keep composing: feed the result into a Chat node to actually run a conversation, or chain group chats.

The hard truth

This is the node where the pack's unfinished state bites hardest. AutoGen group chats are famous for termination problems, and NodeGPT's were no exception - people reported the queue running away until they closed ComfyUI entirely, and agents that echoed each other. The author's own words: "not everything is functioning correctly yet." And the repo is gone: xXAdonesXx/NodeGPT 404s today, author's account deleted, so the install below is history, and the surviving fork only has the pack's earliest build.

cd ComfyUI/custom_nodes
git clone https://github.com/xXAdonesXx/NodeGPT

Restart to auto-install requirements (including pyautogen), wire LLM + User + two agents, set max_round low, and experiment. The pattern is worth studying even now - multi-agent task solving is a live field - but in 2026 you'd build this with a maintained framework, not this fossil.