Go to ACL ARR 2026 January homepageAnemoi: A Semi-Centralized Multi-agent System Based on Agent-to-Agent Communication
Keywords: Large Language Model, Multi-agent System, Agent Communication Protocol
Abstract: Recent advances in generalist Multi-Agent Systems (MASs) have largely converged on a hierarchical fixed workflow paradigm in which a centralized planner agent decomposes tasks and coordinates multiple worker agents via unidirectional prompt passing.
This design has proven effective when powered by strong large-scale language models. However, it exhibits an over-reliance on planner capability, leading to sharp performance degradation in small-LLM settings.
To address this challenge, we propose Anemoi, a semi-centralized dynamic MAS based on Agent-to-Agent (A2A) communication.
Unlike hierarchical fixed workflows, Anemoi does not prescribe a coordination procedure. Instead, coordination emerges from agent-level decisions, allowing agents to monitor progress, assess intermediate results, identify bottlenecks, propose refinements, or disengage when sufficient confidence is reached. A planner provides an initial plan, while A2A enables adaptive task decomposition and collaborative refinement.
Evaluated on the GAIA benchmark, Anemoi achieves 52.73\% accuracy using only small LLMs (GPT-4.1-mini as planner and GPT-4o as worker agents), surpassing the previous open-source state-of-the-art OWL (47.27\%) under identical model and agent settings.
Error and case analyzes further demonstrate that Anemoi’s gains primarily stem from A2A discussion in task decomposition, which effectively mitigates the performance bottleneck of the planner agent.
Our implementation is publicly available at: https://anonymous.4open.science/r/Anemoi-3654.
Paper Type: Long
Research Area: AI/LLM Agents
Research Area Keywords: LLM agents, multi-agent systems, planning in agents, agent communication, agent coordination and negotiation
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 9588
Loading