Multi-agent rooms: the turn budget is per sender and per connector, so one human post can produce N×max_turns messages #188

Summary

agent_chain.max_turns (default 5) is counted per sender, and each connector keeps its own in-memory TurnStore (see #187). Nothing bounds what a room produces for one human message. With 5 agents the ceiling is 5 × 5 = 25 messages.

Observed (2026-10-07, Mattermost, 5 agents in #product-idea)

The owner reported "I write 1 message, I get ~25 messages back, then spend 10–20 minutes catching up." That is exactly N agents × max_turns. The cost is not model quality (the discussion was good) but human reading bandwidth: the owner stopped speaking in the room because they could not keep up, and rejected "let the agents discuss off-screen and summarise" because it loses quality.

Why it matters

In a multi-agent room the human is the slowest participant. The only existing control is a loop-protection cap, which limits runaway chains but is not designed to pace a conversation.

Possible direction (for triage, not a decision)

• A room-level budget: at most N agent messages per human message, shared across the room's bots (this needs the counter to live above the per-connector TurnStore).
• A pacing mode: agents hold after the budget until the human releases the next round, versus free-running with interruption rights. The room discussion argued both are legitimate and the right choice depends on the scenario, so this probably wants to be a per-rule or per-scenario setting.

Related

• Agent-chain counters are not reset by a human post that does not @mention the bot (channel rooms) #187 (human post does not reset counters; per-connector stores)