Anemoi: A Semi-Centralized Multi-agent System Based on Agent-to-Agent Communication MCP server from Coral Protocol
NeurIPS 2025 Workshop on Language Agents and World Models (LAW), 2025
Most LLM multi-agent systems funnel every decision through one central planner. That planner has to hold the entire task in context, which makes it both the performance ceiling and the single point of failure, and it forces you to pay for a large, expensive model in the one role you cannot afford to have fail.
Anemoi takes the opposite position. Worker agents talk to each other directly over an MCP server implementing an agent-to-agent (A2A) communication protocol, so the plan can be monitored and refined by the agents actually executing it rather than dictated up front by a planner that cannot see execution state.
Results on the GAIA benchmark:
- 52.73% accuracy using GPT-4.1-mini as the planner, a deliberately small model.
- +9.09% over the OWL baseline under identical LLM settings.
- Improved scalability and robustness on long-horizon tasks, where centralized planning degrades most.
The finding that matters: the coordination structure, not planner size, was the binding constraint. Distributing the planning burden let a smaller model outperform a centralized system built on a larger one.