Multi-Agent Communication Models: A Comprehensive Guide

Multi-Agent Communication Models: A Comprehensive Guide

Introduction

As artificial intelligence moves from isolated systems to interconnected ecosystems, the ability for AI agents to communicate effectively has become a critical infrastructure challenge [3]. Multi-agent communication models define the rules, message structures, and interaction patterns that enable agents to share information, coordinate decisions, and form shared plans [3].

Modern agent ecosystems require lightweight, modular, cross-platform protocols that integrate smoothly with existing software stacks and support dynamic discovery of agent capabilities [2]. This guide explores the core concepts, classic models, modern protocols, and design principles for multi-agent communication.

What Is Agent Communication?

Agent communication in multi-agent systems is the process of exchanging messages between agents in order to collaborate [3]. Similar to humans, artificial agents use a communication language or protocol to "talk" to one another and to others in their environment [3].

At a high level, effective human communication has three aspects [2]:

  • Communication layer: A channel so both sides can speak and hear each other through a given medium
  • Syntactic layer: A shared syntax so that messages are syntactically understandable
  • Semantic layer: A process for aligning meaning—clarifying questions, confirmations, and repairs that bridge what is said and what is meant

The key to effective multi-agent communication is not just enabling agents to exchange messages, but ensuring they interpret requests, constraints, and outcomes in compatible ways [2].

Classic Communication Models

Early research in multi-agent systems established dedicated languages for inter-agent messaging. These introduced structured message formats and semantics inspired by speech-act theory and have influenced modern protocols [3].

KQML – Knowledge Query and Manipulation Language

KQML was created in the early 1990s as part of DARPA's Knowledge Sharing initiative [3]. KQML defines a set of message performatives (or verbs) to declare the purpose of a message, including actions like ask (query for information), tell (supply information), achieve (request an action), and reply (answer a query) [3].

In KQML, the content of a message is separated from the communication wrappers. A KQML message is a list where the first element is the performative, followed by message parameters [3]:

(ask-one
  :sender Agent1
  :receiver Agent2
  :content "(temperature ?x)"
  :language LPROLOG
  :ontology weather)

KQML also introduced the concept of communication facilitator agents—brokers or mediators that route messages and help agents find each other [3].

FIPA ACL – Foundation for Intelligent Physical Agents

The next major standard following KQML was the FIPA ACL, widely deployed during the late 1990s and early 2000s [3]. FIPA ACL refined the list of performatives and formalized the semantics based on the agents' mental states (beliefs, desires, intentions) [3].

A FIPA ACL message has a fixed set of parameters [6]:

(performative INFORM
:sender Agent1
:receiver Agent2
:ontology WeatherOntology
:language JSON
:content "{ 'forecast': 'sunny' }"
:conversation-id conv123)

Key FIPA performatives include [3]:

  • inform – Inform another agent of some information
  • request – Request that another agent perform some action
  • cfp – Call for proposals (contract-net interaction)
  • propose – Make a proposal in response to a CFP
  • accept-proposal / reject-proposal – Accept or reject a proposal
  • agree – Agree to perform a requested action
  • refuse – Refuse to perform a requested action
  • failure – Indicate an action failed to complete
  • query-if / query-ref – Make a yes/no query or request a specific item
  • subscribe – Request continuous notification
  • not-understood – Signal a message was not understood

Modern Communication Protocols

Classic protocols like KQML and FIPA-ACL emphasized symbolic performatives and logic-based semantics, but their reliance on formal ontologies made them challenging to scale and cumbersome to integrate with web-era APIs [2]. The emergence of LLM-based agents triggered a renewed wave of protocol innovation [2].

Model Context Protocol (MCP)

MCP defines standardized transport channels (stdio, WebSocket, HTTP), capability registration, and structured JSON-based tool invocation [2]. It provides a single standardized interface for connecting to any MCP-compliant tool service, solving the N×M integration problem [11].

Agent-to-Agent (A2A) Protocol

Google's A2A protocol supports cross-platform agent messaging, capability negotiation, and secure agent-level identity management [2]. It enables agents from different frameworks to collaborate through structured task delegation and asynchronous progress updates [11].

Internet of Agents (IoA) Protocol

The IoA Protocol enables dynamic team formation, adaptive task coordination, and structured communication among agents with diverse architectures, tools, and knowledge sources [10]. It features [10]:

  • Agent Registry Block: Stores capability descriptions of all registered agents
  • Team Formation Block: Initiates, joins, or disbands agent teams based on task requirements
  • Session State Machine: Five abstract states: Discussion, Synchronous Task Assignment, Asynchronous Task Assignment, Pause and Trigger, Conclusion
  • Structured message format: Header (sender, group_id) and payload (message_type, next_speaker)

Architectural Communication Patterns

LLM-based multi-agent systems can be organized by communication architecture [7]:

Architecture Description Common Paradigm
Flat All agents communicate as equals [7] One-by-one or simultaneous-talk [7]
Hierarchical Supervisor/manager agents coordinate workers [7] One-by-one with summarizer [7]
Team Agents collaborate as a group with shared goals [7] Simultaneous-talk-with-summarizer [7]
Society Large-scale emergent coordination [7] Simultaneous-talk with environment interaction [7]

Communication Paradigms

Communication paradigms describe how agents exchange messages [7]:

  • One-by-one: Agents speak in sequence; suitable for hierarchical coordination
  • Simultaneous-talk: Multiple agents communicate concurrently; efficient for parallel work
  • Simultaneous-talk-with-summarizer: Concurrent communication with a summarizer agent that consolidates information; balances efficiency with coherence

Recent work on message passing agent systems (MPAS) shows that moving beyond sequential communication to parallel message propagation reduces average communication time from 84.6 seconds to 14.2 seconds per round [8].

Communication Content and Semantics

Agent communication can carry [7]:

  • Explicit communication content: Directly stated information, requests, or instructions
  • Implicit communication content: Contextual cues, inferred intent, or shared understanding

Communication protocols in modern systems increasingly support [2]:

  • Transport, streaming, schema definition, and lifecycle management
  • Limited protocol-level mechanisms for clarification, context alignment, and verification
  • Semantic responsibilities often pushed into prompts or wrappers, creating hidden interoperability costs

Efficiency and Token Optimization

Agentic communication introduces unique efficiency challenges because all exchanged messages are recursively processed by LLMs. Every transmitted token consumes part of the model's limited context window [14].

The Agentic Data Optimization Layer (ADOL) addresses token bloat through [14]:

  • Schema deduplication: JSON $ref references eliminate repetitive definitions
  • Adaptive optional inclusion: Clients can request concise or full schemas
  • Controllable response verbosity: Models specify required output scope
  • Context-aware tool selection: Limit tool exposure to those relevant to the current task

The Semantic Gap in Current Protocols

A detailed analysis of 18 agent communication protocols reveals a clear imbalance in current protocol design [2]:

  • Mature support: Transport, streaming, schema definition, and lifecycle management
  • Limited support: Protocol-level mechanisms for clarification, context alignment, and verification
  • Hidden costs: Semantic responsibilities pushed into prompts, wrappers, or application-specific logic, creating interoperability and maintenance costs [2]

The W3C Semantic Agent Communication Community Group is developing an ontology to define the semantic meaning of agent interactions, anchored to hardware roots of trust through attestation [13].

Related Concepts

  • Multi-Agent Systems — Collaboration, Communication Patterns, Orchestration
  • Tool Calling Fundamentals — The essential concepts and workflow of tool calling
  • API Integration for Agents — Connecting agents to external services
  • External Service Orchestration — Coordinating multiple tools and services
  • Model Context Protocol — Standardized tool connectivity protocol
  • Secure Tool Execution — Protecting agent-tool interactions

Conclusion

Multi-agent communication models are the essential infrastructure for creating an interconnected "Internet of Agents" [2]. The field has evolved from classic speech-act-based protocols like KQML and FIPA ACL to modern LLM-friendly standards like MCP, A2A, and IoA [3] [2].

Key trends include [2] [14]:

  • Standardized protocols for cross-platform agent collaboration
  • Token-efficient communication to reduce context bloat
  • Semantic alignment mechanisms beyond message passing
  • Security, trust, and observability as first-class concerns

For developers building multi-agent systems, understanding communication models is not optional—it is the foundation that determines whether agents can collaborate effectively, scale to complex tasks, and operate reliably across heterogeneous environments.

References

  1. Chen, Jingdi, et al. The Five Ws of Multi-Agent Communication: Who Talks to Whom, When, What, and Why. Transactions on Machine Learning Research. 2026.
  2. Yuan, Dun, et al. Beyond Message Passing: A Semantic View of Agent Communication Protocols. arXiv. 2025.
  3. DigitalOcean. Agent Communication Protocols Explained. DigitalOcean. 2025.
  4. UMBC Department of Computer Science. What is KQML?. UMBC. 2025.
  5. Charles University. Communication Protocols. Department of Theoretical Computer Science. 2025.
  6. Yan, Bingyu, et al. Beyond Self-Talk: A Communication-Centric Survey of LLM-Based Multi-Agent Systems. arXiv. 2025.
  7. Yu, Jingxuan, et al. MPAS: Breaking Sequential Constraints of Multi-Agent Communication Topologies via Individual-Epistemic Message Propagation. AAAI. 2026.
  8. IETF. Framework, Use Cases and Requirements for AI Agent Protocols. Internet-Draft. 2025.
  9. Yang, C., et al. Internet of Agents Protocol (IoA Protocol) for Heterogeneous Agent Collaboration. IETF. 2025.
  10. Chang, Z., et al. A Token-efficient Data Layer for Agentic Communication. IETF. 2025.
  11. W3C Semantic Agent Communication CG. Agent Semantic Communication Ontology. GitHub. 2025.
  12. IEEE. AI-Agent Communication Network for 6G. IEEE Xplore. 2025.
  13. AgentGateway. AgentGateway Documentation. Go Packages. 2025.

Comments

Post a Comment