Creativity · Agent Protocol
Agent Voting / Ensemble Pattern
For high-stakes or ambiguous tasks, running one agent is gambling. The voting-ensemble pattern spawns N agents (same model with different seeds, or a mix of models) on the same prompt and picks the answer by majority vote or uses a judge LLM to pick the best. You pay Nx the cost in exchange for dramatically lower tail-risk errors.
Protocol facts
- Sponsor
- Research community
- Status
- stable
- Interop with
- LangGraph, AutoGen, any parallel agent framework
Frequently asked questions
What's the difference between voting and judge-based aggregation?
Voting counts identical answers — works for multiple-choice or extractable facts. Judge-based uses an LLM to read all N responses and pick/synthesize the best — works for open-ended work like essays or code.
Should the N agents be identical?
Not necessarily. Diverse ensembles (different models, different prompts, different temperatures) often outperform N copies of the same agent, because diverse mistakes cancel out.
When is it not worth it?
For simple, well-bounded tasks where a single strong agent is already >99% accurate. The extra cost buys nothing, and latency increases linearly with N unless you parallelize.
Sources
- Self-Consistency paper (Wang et al. 2022) — accessed 2026-04-20
- LLM-as-a-Judge paper — accessed 2026-04-20