When two AI agents built by different companies need to collaborate on a task, they face a problem that would be familiar to anyone who remembers the early days of computer networking: they do not speak the same language. The current ecosystem of LLM-powered agents relies on ad-hoc integrations — brittle, vendor-specific glue code that breaks every time an API changes. This fragmentation is not merely inconvenient. It poses a significant barrier to realizing the promise of autonomous multi-agent systems.
The Fragmentation Problem
The situation echoes what networking engineers faced in the 1970s and 1980s. Before TCP/IP emerged as the universal standard, competing protocols from IBM, Xerox, and others created isolated islands of connectivity. Li, Liu, and Yuen (2025) draw this parallel explicitly, warning that without a principled standardization effort, the current path of fragmentation will ultimately limit the potential of this transformative technology.
Today, at least four major agent communication protocols are competing for adoption, each backed by a different technology giant or community:
- Model Context Protocol (MCP), launched by Anthropic in November 2024, provides a JSON-RPC client-server interface for structured tool invocation and context delivery — essentially standardizing how applications feed tools, datasets, and instructions to LLMs.
- Agent-to-Agent Protocol (A2A), released by Google in April 2025, enables peer-to-peer collaboration using capability-based "Agent Cards" that describe what each agent can do and how it can be reached.
- Agent Communication Protocol (ACP), introduced by IBM under the Linux Foundation in March 2025, offers a RESTful, SDK-optional interface with role-based access control and session management.
- Agent Network Protocol (ANP), a community-driven open standard, uses W3C Decentralized Identifiers (DIDs) and JSON-LD to enable agent discovery and secure collaboration across the open internet.
Four Protocols, Four Design Philosophies
What makes this competition intellectually interesting is that each protocol targets a different tier of the interoperability stack. Ehtesham, Singh, Gupta, and Kumar (2025) provide the most comprehensive comparative analysis to date, examining these four protocols across multiple dimensions including interaction modes, discovery mechanisms, communication patterns, and security models.
MCP solves the "context delivery" problem — how to get the right tools and data to an LLM in a standardized way. The survey describes it as akin to a USB-C for AI, providing a universal, model-agnostic interface. But MCP operates in a client-server paradigm; it does not address how two autonomous agents should discover each other or negotiate tasks.
A2A fills that gap. Its Agent Cards function as machine-readable capability advertisements, enabling agents to find and invoke each other without prior configuration. It supports asynchronous, event-driven communication through HTTP and Server-Sent Events, making it suitable for distributed, scalable agent ecosystems.
ACP takes a different approach, prioritizing enterprise-grade features: structured session management, message routing, and flexible authentication including RBAC and decentralized identity integration. Its REST-based design makes it compatible with existing enterprise infrastructure.
ANP pushes furthest toward decentralization. By adopting a peer-to-peer model with W3C DIDs, it allows agents to autonomously discover, authenticate, and interact without centralized intermediaries — the most ambitious vision, but also the most challenging to deploy.
A Phased Adoption Roadmap
Rather than declaring a winner, Ehtesham et al. propose a pragmatic, phased adoption roadmap: organizations should start with MCP for tool access, layer in ACP for structured multimodal messaging, add A2A for collaborative task execution, and eventually extend to ANP for decentralized agent marketplaces.
This layered strategy mirrors how the internet itself evolved — TCP/IP did not replace everything overnight. It became the narrow waist that allowed innovation above and below it. The question for the AI agent ecosystem is whether any of these protocols can achieve that same universal adoption, or whether we are destined for a prolonged period of competing standards.
The Telecom Lesson
Li et al. (2025) argue that the telecommunications industry offers a direct blueprint. The success of the OSI model and later TCP/IP depended on three principles: consensus-driven open standards through bodies like ITU and 3GPP, security by construction rather than as an afterthought, and layered abstractions that allow different parts of the stack to evolve independently.
Their proposed LACP (LLM-Agent Communication Protocol) embodies these principles in a three-layer architecture: a Semantic Layer for conveying intent through universal message types (PLAN, ACT, OBSERVE), a Transactional Layer for ensuring reliability through message signing and atomic transactions, and a Transport Layer that is agnostic to the underlying network protocol.
Open Questions
Several critical tensions remain unresolved. Can a single protocol serve both the controlled enterprise environment where ACP excels and the open internet where ANP is designed to operate? Will the major players — Anthropic, Google, IBM — converge on a shared standard, or will market dynamics produce a VHS-versus-Betamax outcome? And perhaps most importantly, how do we build security into agent communication protocols from the ground up, rather than bolting it on after deployment?
The history of networking suggests that standardization is inevitable — the question is whether it happens through deliberate collaboration or through the slow, painful process of one protocol achieving dominance by network effects alone.
Looking Forward
The next twelve months will be decisive. As multi-agent systems move from research prototypes to production deployments in healthcare, finance, and autonomous systems, the cost of interoperability failures will become impossible to ignore. The organizations and communities that shape these protocols today will determine the architecture of the agent economy for years to come.
References
Ehtesham, A., Singh, A., Gupta, G. K., & Kumar, S. (2025). A survey of agent interoperability protocols: Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent-to-Agent Protocol (A2A), and Agent Network Protocol (ANP). arXiv preprint, arXiv:2505.02279.
Li, X., Liu, M., & Yuen, C. (2025). LLM Agent Communication Protocol (LACP) requires urgent standardization: A telecom-inspired protocol is necessary. NeurIPS 2025 AI4NextG Workshop. arXiv:2510.13821.