Multi-agent collaboration: blind rounds and convergence

How to structure multi-agent brainstorming so agents produce independent ideas before they see each other's work — and what to do when they converge anyway.

TL;DR — Multi-agent brainstorming only adds value if the agents are independent first and social second. Run a blind round for divergence, then an open round for combination. When round two collapses to consensus, that is data about the brief, not a failure.

Why independence is the whole game

If you give several agents the same prompt in the same shared context, you do not get several ideas — you get one idea with several voices. Shared context creates anchoring, and anchoring destroys the reason to run more than one agent.

The structure that works

Phase 1 — blind divergence

• Each agent gets the same brief, no shared scratchpad.
• No agent sees another's output until all have submitted.
• Score each idea for independent novelty (R1) — how unlikely it was to arise from the default path (see splitting novelty).

Phase 2 — open response

• Every agent now reads the full set.
• Each must name the specific idea it builds on, rebuts or combines.
• Score for incremental novelty (R2) — the value added over what already exists.

The convergence signal

A common worry: two rounds of the same model converge, and round two adds little. Two readings:

1. Bad process — agents saw each other too early, so round one was never independent.
2. Narrow brief — the brief had one obvious answer, so both rounds found it. That is a finding about the problem, not the process.

Distinguish them by checking R1 spread: if round one was genuinely diverse and round two still converged, the brief is the constraint.

Practical rules

• Never share context across round one. Independence is fragile; a single example anchors everyone.
• Require explicit references in round two. "Building on idea 7 by adding X" — vague responses get rejected.
• Keep the human call. Convergence is information, not a verdict; a human decides whether to widen the brief or accept the consensus.

This is the mechanism the framework automates. See two-round brainstorm for the manual version.