AI Agent Protocols Are Proliferating—Here’s What Engineering Leaders Need to Know
The AI agent ecosystem is experiencing a surge in protocol development, mirroring the history of distributed computing. Four major standards—Model Context Protocol (MCP), Agent Communication Protocol (ACP), Agent2Agent (A2A), and Agent Network Protocol (ANP)—have emerged in the past 18 months, each addressing distinct layers of agent communication. While these protocols are not direct competitors, their coexistence creates complexity for developers. Understanding their roles and the unresolved challenges of peer-to-peer networking is critical for architecture decisions today.
What Each Protocol Solves
MCP, developed by Anthropic, functions as a tool-calling interface, enabling models to discover, invoke, and interpret functions from servers. By April 2026, the Linux Foundation reported over 10,000 active public MCP servers and 164 million monthly Python SDK downloads, cementing its dominance in the tool-calling layer. A2A, created by Google, focuses on task coordination between agents, introducing capabilities like Agent Cards and task lifecycle states. It has been widely adopted by enterprise AI teams to fill gaps left by MCP.
ACP, an IBM Research initiative, serves as a lightweight message envelope format for agent-to-agent communication, avoiding the overhead of A2A’s task coordination semantics. ANP, developed by an independent working group, handles agent discovery and identity using Decentralized Identifiers (DIDs) and JSON-LD graphs, enabling decentralized agent marketplaces. Together, these protocols form a stack: ANP for capability discovery, A2A for task coordination, MCP for tool calls, and ACP for simple messaging.
The Transport Layer Challenge
Despite progress in application-layer protocols, the transport layer remains unresolved. All current protocols operate over HTTP, a legacy choice driven by familiarity and ease of demonstration. However, HTTP’s requirement for a reachable server creates bottlenecks in environments with Network Address Translation (NAT), which affects 88% of devices. This centralization forces messages through relay infrastructure, adding latency and cost.
Technologies like UDP hole-punching with STUN, X25519 Diffie-Hellman encryption, and QUIC (RFC 9000) offer solutions for direct peer-to-peer communication. However, these tools must be adapted for capability-based routing, where agents query peers based on their functions rather than IP addresses. Projects like Pilot Protocol and libp2p are exploring these challenges, with the IETF’s QUIC working group developing NAT traversal extensions.
What the Future Holds
Application-layer protocols like MCP and A2A are nearing stability, with expected refinements in security and scalability. The transport layer, however, faces 18–24 months of experimentation before consolidation. The W3C and IETF are expected to publish formal standards between 2027 and 2028, but open-source implementations may establish de facto norms earlier.
For engineering leaders, the key takeaway is layered adoption. Application-layer protocols are production-ready, while transport solutions remain experimental. Systems designed with a clear separation between application semantics (MCP, A2A) and transport (e.g., QUIC, custom UDP protocols) will be better positioned to adapt as standards evolve.
Why This Matters
The evolution of AI agent protocols reflects broader trends in technology: initial fragmentation followed by gradual convergence. Historically, protocols like REST supplanted earlier standards by prioritizing simplicity and interoperability. The same dynamic is playing out in the AI agent space, with HTTP-based protocols already gaining traction. However, the transport layer’s reliance on NAT traversal highlights the need for innovation in decentralized networking—a challenge that could shape the next generation of AI systems.