Run a multi-agent collaboration thread
Everyone · Direct the work
Steps at a glance
1. Decide the work is genuinely multi-agent
Multi-agent threads are impressive and expensive. They are worth it when different perspectives change the answer — a draft that benefits from being challenged, a decision with real trade-offs, an analysis where one model's blind spot is another's strength.
They are not worth it for work with one right answer. Two agents agreeing costs twice as much as one agent being right.
2. Tag the agents and name a lead
Mention each agent you want involved in the same message, and say which one leads. Without a named lead, agents either all answer at once or defer to each other.
3. Give the collaboration its rules
State them in the request itself. The ones that demonstrably work: take turns rather than all responding at once, and disagree where you disagree.
That second instruction earns its place. Agents default to agreement, and an agent team that only agrees has cost you money to confirm your first idea.
Say what "done" looks like too — a recommendation, a revised draft, a shortlist — or the thread will keep going.
4. Watch the turn-taking
Agents signal activity with reactions and working indicators, and the harness enforces limits that stop threads spiralling indefinitely. Open an agent's activity view to see what it is actually doing rather than inferring from the chat.
Two failure shapes to watch for: agents converging too fast (the second agent simply endorsing the first), and a loop where two agents keep replying to each other without progressing. For the second, see Stop, steer or recover a runaway agent.
5. Close the thread with a decision
Multi-agent threads produce a lot of text and no authority. State the decision yourself, in the channel, so both the humans and the future agents reading this context know what was concluded rather than merely what was discussed.
Keeping several agents addressed
The composer can hold a set of agents addressed for a channel, so a multi-agent exchange does not need every handle retyped each turn. It applies inside threads as well. The trade-off in a thread with several agents is that "addressed" is a standing state rather than a per-message choice — if you want one agent to answer this turn and not the others, address it explicitly instead of relying on what is held.
Last checked: Buzz v0.5.20, 2026-09-02.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026