Multi-Agent Systems

Multi-agent systems represent the third layer of the agentic reasoning framework — extending intelligence from individual agents to collaborative settings where multiple LLM-powered agents coordinate to solve problems. This area is rapidly evolving as standardized protocols emerge for agent-to-agent communication.

Ferrag et al. survey three key collaboration protocols that are shaping how agents interoperate: ACP (Agent Collaboration Protocol), MCP (Model Context Protocol), and A2A (Agent-to-Agent). These protocols define how agents discover each other's capabilities, negotiate task delegation, and share results — moving from ad-hoc multi-agent setups to standardized infrastructure.

Wei et al. frame multi-agent reasoning as "collective intelligence" — the highest layer of agentic capability, where the challenges shift from individual reasoning to coordination, negotiation, and emergent group behavior.

Key Claims

• Three collaboration protocols are standardizing agent interoperability — ACP, MCP, and A2A define discovery, delegation, and result-sharing between agents. Evidence: strong (From LLM Reasoning to Autonomous Agents)
• Multi-agent is the highest layer of agentic reasoning — Extends foundational and self-evolving capabilities to collaborative settings. Evidence: strong (Agentic Reasoning for LLMs)
• Governance of multi-agent systems is an open problem — Safety, alignment, and accountability become harder with multiple autonomous agents. Evidence: strong (Agentic Reasoning for LLMs)

Open Questions

• How to ensure safety when multiple autonomous agents interact without human oversight?
• Can standardized protocols (MCP, A2A) scale to thousands of heterogeneous agents?
• How to attribute responsibility when a multi-agent system produces harmful outputs?
• What coordination mechanisms prevent emergent adversarial dynamics between agents?

Related Concepts

• Agentic Reasoning — Multi-agent is the third layer of the framework
• LLM Tool Use — Agents use tools as part of collaborative task execution
• Agent Evaluation Benchmarks — Interactive benchmarks increasingly test multi-agent scenarios

Backlinks

Pages that reference this concept:

• OpenAI