Anemoi

renxinxing123/Anemoi

3.2

If you are the rightful owner of Anemoi and would like to certify it and/or have it hosted online, please leave a comment on the right or send an email to henry@mcphub.com.

Anemoi is a semi-centralized multi-agent system (MAS) that leverages Agent-to-Agent (A2A) communication on an MCP server to enable efficient and scalable inter-agent collaboration.

🌬️ Anemoi: A Semi-Centralized Multi-Agent System Based on Agent-to-Agent Communication MCP Server

Anemoi is a semi-centralized multi-agent system (MAS) built on the Agent-to-Agent (A2A) communication MCP server. Unlike traditional context-engineering + centralized paradigms, Anemoi enables structured and direct inter-agent collaboration, allowing agents to debate, refine, and adapt in real time.

✨ Motivation

Recent advances in generalist MAS largely follow a centralized planner + worker paradigm, where:

  • A planner agent coordinates multiple worker agents through unidirectional prompt passing.

  • This design works with strong LLMs, but faces two critical limitations:

    1. Over-reliance on large LLMs → performance drops sharply in small-LLM settings.
    2. Limited inter-agent communication → collaboration is reduced to prompt concatenation, not genuine refinement.

Anemoi addresses these challenges by introducing semi-centralized A2A collaboration with Multi-Agents Debate, making agents behave more like a real-world team.

🚀 Key Features

  • Semi-Centralized Architecture Reduces dependency on a single planner agent, enabling adaptive updates.

  • Multi-Agents Debate for Collaboration All agents can monitor progress, assess results, identify bottlenecks, and refine plans in real time.

  • Collaboration-Driven Reliability Multi-Agents Debate yields more consistent and explainable outcomes than stochastic worker behavior.

📊 Benchmark Results (GAIA)

We evaluated Anemoi on the General Artificial Intelligence Assistants (GAIA) benchmark — a challenging suite of real-world, multi-step tasks (web search, file processing, coding).

SettingPlannerWorkersAccuracy (pass@3)Comparison vs. OWL
Small-LLM (SOTA)GPT-4.1-miniGPT-4.1-mini / o4-mini63.64%+1.82%
Weaker WorkersGPT-4.1-miniGPT-4o52.73%+9.09%

📌 Highlights:

  • Anemoi establishes a new SOTA in the small-LLM regime.

  • Outperforms Optimized Workforce Learning (OWL) under identical settings.

  • Case-level analysis:

    • 25 tasks solved by Anemoi but missed by OWL → 48% enabled by Multi-Agents Debate.
    • OWL’s unique wins (86%) mainly reflect stochastic worker behavior.

Reproduction

Set up environment variables:

echo '
export FIRECRAWL_API_KEY="your_firecrawl_api_key"
export GOOGLE_API_KEY="your_google_api_key"
export HF_HOME="your_hf_home_path"
export AZURE_KEY="your_azure_api_key"
export SEARCH_ENGINE_ID="your_search_engine_id"
export CHUNKR_API_KEY="your_chunkr_api_key"
' >> ~/.bashrc && source ~/.bashrc

Create environment:

cd Anemoi
/usr/bin/python3.12 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

We made some minor modifications to CAMEL 0.2.70 for our experiments:

rm -rf venv/lib/python3.12/site-packages/camel
cp -r utils/camel venv/lib/python3.12/site-packages/

Run the experiment:

cd ..
./gradlew run --console=plain