Consensus voting

Use this when a single AI judge isn't enough: the cost of being wrong is high enough that you want N independent agents to vote, and you only proceed when K of them agree. Think refund exceptions, content moderation, code-review approval, anything where a one-shot model call is too unreliable.

The workflow#

A consensus node fans out the same prompt to agentCount agents in parallel and routes to approved only when at least quorum of them vote yes. Errored or timed-out agents count as no-votes: they neither approve nor block, and they don't count toward the quorum at all.

name: consensus
version: 1
description: |
  Multi-agent gate where 3 independent reviewers must reach a 2-of-3 quorum
  before the workflow proceeds.
initial: ask
states:
  ask:
    type: consensus
    label: Is this idea worth shipping?
    model: claude-haiku-4-5
    prompt: |
      Evaluate this product proposal:

        "Add a feature where users can attach short voice memos (up to 30s) to
         any item in their task list, transcribed automatically and searchable
         alongside text notes."

      Approve only if all four hold:
        1. it solves a real, common user problem
        2. it can ship in under 4 weeks of engineering
        3. it doesn't add meaningful privacy/compliance burden
        4. it's not better delivered as an integration

      Be skeptical — reject if any single criterion is shaky. Answer with
      the JSON shape requested below.
    agentCount: 3
    quorum: 2
    validator:
      kind: boolean
    maxRetries: 1
    on:
      approved: yes-end
      rejected: no-end

  yes-end:
    type: end
    label: Approved
    outcome: success
    message: Reviewer quorum agreed.

  no-end:
    type: end
    label: Rejected
    outcome: failure
    message: Reviewer quorum did not agree.

Running it#

roscoe run consensus

The web UI shows a verdict card with each agent's vote and reasoning. When agents split, you'll see a "split decision" banner and per-agent reasoning panels. That's the whole point of the node, so design prompts that genuinely let models disagree.

Note: consensus always calls the LLM directly, even in MCP mode. Unlike ai_agent and ai_judge, it does not pause for Claude Code to step through each agent: doing so would defeat the "atomic quorum" guarantee. Running this workflow needs either ANTHROPIC_API_KEY set or the claude CLI on PATH.

Tuning agentCount and quorum#

`agentCount`	`quorum`	Behaviour
3	2	Default. Two-of-three majority. Cheap; tolerant of one outlier.
3	3	Unanimous. Any dissent rejects. Use when false positives are very expensive.
5	3	Bare majority of five. More signal, ~1.7x the cost of 3-of-3.
5	4	Strong majority. Tolerant of one outlier; rejects on two.
7	5	High-stakes, expensive. Rare in practice.

agentCount is bounded 2..7. quorum must be ≤ agentCount: the schema rejects quorum > agentCount outright, since such a node could never approve.

If quorum equals agentCount you've made the gate strictly unanimous; one errored agent is enough to block. If you care more about latency than safety, prefer odd agentCount and quorum = ceil(agentCount/2).

When to reach for it#

For a refund exception, three agents check the request against your refund policy and approve only if 2 of 3 agree it's a genuine edge case. For content moderation, five agents check a piece of user content and publish only if 4 of 5 say it's safe. For code review approval, three agents review a diff against the same rubric and ship only if 2 of 3 approve.

For lower-stakes branching (single-agent fact checks, feature toggles, light guardrails), ai_judge is the right call. consensus is several times the cost.

Variations#

For a strictly unanimous variant, set quorum: 3 with agentCount: 3; any dissent rejects. To vary the question per run, embed a {{ field }} reference in the prompt (declared under the workflow's inputs:) and pass its value via roscoe run consensus --var field=value. To chain into a human gate, route approved to a human node titled "Final sign-off" instead of straight to an end; useful when you want AI to filter the obvious cases but a human to confirm the edge ones.