Agent Communication Protocols: An Academic Survey

Understanding the Emerging Landscape of AI Agent Interoperability

Document ID: STCMastery-ACP-daee8f3e-e23b-4c9a-947d-2bc92abf3c3f
Date: February 7, 2026

Executive Summary

Agent Communication Protocol (ACP), alongside Model Context Protocol (MCP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP), represents the current generation of standardized communication frameworks enabling LLM-powered agents to discover, coordinate, and collaborate across heterogeneous systems. This survey synthesizes recent academic literature to provide an accessible introduction to the state of the field, current research inquiries, and practical implementation considerations for multi-agent system developers.

1. The Interoperability Crisis and Protocol Emergence

The proliferation of LLM-based autonomous agents has created a fundamental interoperability crisis[cite:1][cite:2]. Developers building agents with frameworks like LangChain, CrewAI, BeeAI, and custom runtimes face incompatible messaging formats, orchestration assumptions, and lifecycle patterns. This fragmentation introduces exponential integration costs as agent populations grow, hindering the formation of collective intelligence systems[cite:3][cite:7].

Four standardized protocols emerged in 2024-2025 to address distinct layers of agent interoperability:

Model Context Protocol (MCP) - Created by Anthropic in November 2024 and transferred to the Linux Foundation Agentic AI Foundation in December 2025, MCP provides a JSON-RPC client-server interface connecting agents to external tools and data sources[cite:2][cite:8]. It standardizes how applications deliver Resources, Tools, Prompts, and Sampling capabilities to LLMs, functioning as "USB-C for AI"[cite:2].

Agent Communication Protocol (ACP) - Proposed by IBM Research and BeeAI, ACP is a REST-native protocol optimized for local-first, low-latency agent-to-agent communication[cite:3][cite:7]. Built on RESTful architecture over HTTP with JSON-RPC 2.0, it supports both synchronous and asynchronous interactions, multimodal message exchange via MIME-type extensibility, and offline capability discovery through metadata embedded in distribution packages[cite:3][cite:5].

Agent-to-Agent Protocol (A2A) - Driven by Google with over 50 industry partners including Microsoft, Salesforce, Oracle, and SAP, A2A enables peer-to-peer task outsourcing through capability-based Agent Cards[cite:2][cite:9]. It facilitates enterprise-scale workflows with standardized discovery, authentication, and collaboration mechanisms[cite:46].

Agent Network Protocol (ANP) - ANP supports decentralized agent discovery and collaboration across the open internet using W3C Decentralized Identifiers (DIDs) and JSON-LD semantic graphs[cite:2]. It targets public agent marketplaces and cross-organizational coordination scenarios.

2. Classification Framework and Architectural Patterns

Recent literature proposes a two-dimensional classification distinguishing context-oriented versus inter-agent protocols and general-purpose versus domain-specific protocols[cite:1]. Context-oriented protocols (MCP) focus on enriching agent execution with external knowledge and tools, while inter-agent protocols (ACP, A2A, ANP) enable coordination between autonomous entities.

Architectural Comparison[cite:2]:

MCP follows a three-phase lifecycle: Initialization (version negotiation, capability exchange), Operation (JSON-RPC method calls, timeout management), and Shutdown (resource cleanup)[cite:2]. ACP introduces an asynchronous-first model with multipart MIME messages supporting text, structured JSON, files, and multimedia streams[cite:3][cite:4].

A2A's Agent Card mechanism—a JSON document served at /.well-known/agent.json—declares metadata including name, version, supported skills, and authentication schemes[cite:2][cite:5]. This enables dynamic discovery without live service calls, supporting air-gapped and scale-to-zero environments[cite:3].

3. Current Research Inquiries

Academic literature identifies seven critical research directions shaping next-generation protocols:

3.1 Semantic Interoperability
While protocols standardize message envelopes and transport mechanisms, agents with heterogeneous ontologies struggle to interpret concrete concepts. Recent work proposes combining Semantic Web technology (for abstract concepts) with deep learning approaches (for images and multimodal data) to extend interoperability beyond symbolic reasoning[cite:48]. This builds on historical FIPA Agent Communication Language (FIPA-ACL) efforts from 2000, which used formal semantics grounded in speech act theory and agent mental states (beliefs, desires, intentions)[cite:47][cite:52].

3.2 Communication Efficiency and Goal-Oriented Messaging
Resource-constrained environments—edge devices, IoT systems, swarm robotics—require protocols prioritizing task relevance over message fidelity[cite:36]. Goal-oriented communication frameworks use information theory and machine learning to determine when communication is necessary and what information maximizes collaborative decision-making[cite:36][cite:42]. Attentional communication models enable agents to learn dynamic coordination strategies, reducing unnecessary transmissions while maintaining performance[cite:42].

3.3 Security, Trust, and Observability
Security vulnerabilities span protocol lifecycles: installer spoofing during creation, tool poisoning and credential theft during operation, and privilege persistence across updates[cite:2]. Mitigation strategies include Software Bills of Materials (SBOMs), digital signatures, mutual TLS authentication, OAuth 2.1 with PKCE, and runtime monitoring via EDR/HIDS systems[cite:2][cite:32]. ACP emphasizes enterprise-grade security with API gateway integration (Apigee, Kong), centralized audit trails, and OAuth2/mTLS support[cite:32].

3.4 Scalability and Collective Intelligence
Protocols must support layered architectures enabling hierarchical coordination patterns. The Router Agent pattern—where a central agent decomposes complex requests, routes tasks to specialists, and aggregates responses—demonstrates how protocols can enable modular, scalable workflows[cite:5][cite:50]. Multi-agent orchestration frameworks like Semantic Kernel, CrewAI, and AG2 leverage these patterns for enterprise deployments[cite:2][cite:50].

3.5 Phased Adoption and Protocol Coexistence
Research proposes a four-stage adoption roadmap[cite:2]:

• Stage 1: MCP for tool invocation (agent ↔ APIs, databases)
• Stage 2: ACP for multimodal messaging (local agent coordination)
• Stage 3: A2A for enterprise collaboration (cross-framework task delegation)
• Stage 4: ANP for decentralized markets (global agent discovery)

This layered approach recognizes that MCP and ACP/A2A are complementary rather than competing—MCP connects agents to tools, while ACP/A2A connect agents to each other[cite:7][cite:50].

3.6 Dynamic Discovery and Capability Negotiation
Static tool registries force agent re-initialization when APIs change, preventing truly adaptive systems[cite:2]. Protocols addressing this limitation enable runtime capability advertisement, versioned manifests with digital signatures, and scoped namespaces preventing cross-server shadowing attacks[cite:2]. ACP's offline discovery—where capability metadata is embedded in distribution packages—supports secure, air-gapped environments[cite:3].

3.7 Standardization and Governance
The W3C AI Agent Protocol Community Group is developing official web standards for agent communication, with specifications expected in 2026-2027[cite:51]. The Linux Foundation's Agentic AI Foundation (December 2025) provides governance for MCP development[cite:51][cite:57]. Industry coalitions like AGNTCY (Cisco, LangChain, LlamaIndex) and the 50+ partners supporting A2A signal movement toward "HTTP for agents"—a universal standard enabling internet-scale interoperability[cite:6][cite:9].

4. Experimental Implementations: --experimental-acp

Two major AI platforms expose ACP capabilities through experimental flags:

Gemini CLI: gemini --experimental-acp starts the agent in ACP mode, supporting stdio and TCP transports[cite:16][cite:28]. Configuration via ~/.jetbrains/acp.json enables integration with IDEs like IntelliJ IDEA[cite:16]. Known limitations include proxy flag compatibility and SSE-based MCP transport support[cite:19][cite:25].

GitHub Copilot CLI: copilot --acp --stdio or copilot --acp --port 8080 launches Copilot as an ACP server[cite:17][cite:23][cite:26]. Clients can initialize connections, create isolated sessions, send prompts with multimodal context, receive streaming updates, and respond to permission requests for tool execution[cite:23]. The TypeScript SDK @agentclientprotocol/sdk provides reference implementations for programmatic interaction[cite:17][cite:37].

These experimental flags represent first-generation deployments, validating protocol feasibility while exposing real-world integration challenges (authentication flows, MCP transport compatibility, proxy configurations)[cite:19][cite:22][cite:25].

5. Conclusion and Future Outlook

The agent protocol landscape is consolidating around layered, complementary standards addressing distinct interoperability tiers. Near-term developments (2025-2026) will focus on protocol convergence, W3C standardization, and enterprise tooling maturation[cite:51]. Mid-term evolution (2026-2027) will produce domain-specific variants for healthcare, finance, and manufacturing, alongside lightweight protocols for edge computing[cite:51]. Long-term trajectories (2027-2030) point toward a dominant universal standard, standardized agent marketplaces, and regulatory frameworks for critical infrastructure[cite:51].

For practitioners building multi-agent systems, the current state of research suggests:

1. Start with MCP for tool integration and external data access
2. Adopt ACP for local, low-latency agent coordination
3. Evaluate A2A for cross-framework enterprise workflows
4. Monitor ANP for future decentralized collaboration scenarios

The transition from fragmented, framework-specific integrations to protocol-driven interoperability marks a pivotal moment in AI agent development. As these protocols mature, they enable the collective intelligence systems originally envisioned by multi-agent systems research—agents that discover, negotiate, and collaborate autonomously to solve problems beyond individual capabilities.

References

Academic papers, industry specifications, and technical documentation are cited inline throughout this document. Key sources include:

• Yang et al. (2025). "A Survey of AI Agent Protocols" (arXiv:2504.16736)[cite:1]
• Ehtesham et al. (2025). "Survey of Agent Interoperability Protocols" (arXiv:2505.02279)[cite:2]
• Yan et al. (2025). "Beyond Self-Talk: Communication-Centric Survey" (arXiv:2502.14321)[cite:33]
• Charalambous et al. (2025). "Goal-Oriented Communication in MAS" (arXiv:2508.07720)[cite:36]
• IBM Research. "Agent Communication Protocol" (research.ibm.com)[cite:7][cite:11]
• Official documentation: agentclientprotocol.dev, modelcontextprotocol.io[cite:5][cite:8]
• GitHub repositories: modelcontextprotocol, agentclientprotocol, i-am-bee/acp[cite:18][cite:37][cite:43]

Document prepared for: Local agent educational purposes
Intended audience: Multi-agent system developers, AI researchers, protocol implementers
Next steps: Consult practical implementation guide and GitHub resources for hands-on experimentation