agensus

AI Agent Consensus framework - reduces bias and improves accuracy in multi-agent LLM systems by leveraging multiple perspectives to reach reliable consensus.

Why Consensus? Single LLMs can hallucinate, produce biased results, or generate false information. Agensus mitigates these issues by combining outputs from multiple agents using proven consensus algorithms.

• Strategies: overlap (Jaccard similarity), rrf (reciprocal rank fusion), llm_judge (meta-evaluation)
• API: Consensus(strategy=...).pick(candidates) → ConsensusResult
• Production-ready: Fast, dependency-light, pure Python
• Flexible: CLI + Python API with LLM integration examples

Install

pip install -e .

Quick Start

from agensus import Consensus

# Multiple AI responses to the same question
candidates = [
    "Deploy using blue-green strategy with load balancer switching",
    "Use rolling deployment with health checks and automatic rollback", 
    "Implement canary deployment with gradual traffic shifting"
]

# Find consensus using overlap similarity
result = Consensus("overlap").pick(candidates)
print(f"Best answer: {candidates[result.index]}")
print(f"Confidence scores: {result.scores}")

Advanced Multi-Agent Example

See examples/llm_agents.py for full script pulling multiple model outputs and applying overlap and LLM-judge consensus.

from agensus import Consensus
from examples.llm_agents import generate_candidates, llm_judge_builder

prompt = "Outline a zero-downtime deployment pipeline for a SaaS platform."
models = ["gpt-4o-mini", "gpt-4o"]
answers = generate_candidates(prompt, models)

# Overlap strategy
res_overlap = Consensus("overlap").pick(answers)
print("Overlap winner index:", res_overlap.index)
print("Scores:", res_overlap.scores)

# LLM judge strategy
judge_fn = llm_judge_builder("gpt-4o")
res_judge = Consensus("llm_judge", judge_fn=judge_fn).pick(answers)
print("LLM Judge winner index:", res_judge.index)
print("Rationale:", res_judge.rationale)

Research Background

Based on research showing that multi-agent consensus frameworks significantly reduce bias and improve accuracy in generative AI systems. Single LLMs are prone to hallucinations and biased outputs - agensus provides practical consensus mechanisms to address these reliability challenges.

CLI

printf "A\nB more info\nC longest option" | agensus --strategy overlap

Tests

pytest -q

See docs/ for deeper strategy notes.