AGENT ATLASa field guide to agent workflows
Back to Multi-agent
Multi-agent

Voting / ensemble

Run the same task several times, then combine the answers.

N agentsCore

How it works

  1. 1Run the same prompt multiple times, or with varied framings.
  2. 2Collect the independent answers.
  3. 3Aggregate by majority vote, a threshold, or a judge that reconciles them.
  4. 4Treat disagreement as a signal to escalate to a human.

Use it when

You want reliability on a judgment call — does this code have a bug? is this content safe? — where independent attempts reduce variance.

Reach for something else when

There's one deterministic right answer you can cheaply check, or the cost of N runs isn't justified.

Where you stay in the loop

The ensemble votes; you decide what disagreement means. Treat a split vote as a signal to step in, not a number to average away.

In the wild

Three independent reviewers flag a pull request; merge only if at least two agree it's safe.

Hand this to your agent

Use a voting ensemble.

Answer my question 3 separate times, each as an independent attempt — vary
your reasoning approach each time, don't just restate. Label them Attempt 1–3.

Then aggregate: state the majority answer, note any disagreement, and give a
confidence level. If the three disagree badly, say so and recommend human
review.

Question: <...>

Replace the <…> placeholders, paste it into your agent, and it'll scaffold the workflow with you.