The Evolution of Distributed Computing: What Lies Ahead for AI Agent Protocols

Text
The AI agent ecosystem is in a phase of protocol proliferation, with four major standards emerging in the past 18 months. Model context protocol (MCP) from Anthropic, agent communication protocol (ACP) from IBM Research, Agent2Agent (A2A) from Google, and agent network protocol (ANP) from an independent working group have each addressed distinct layers of agent communication, according to a 2026 analysis by VentureBeat. The W3C AI Agent Protocol Community Group and the Internet Engineering Task Force (IETF) are now formalizing these efforts, with the IETF receiving Internet-Drafts on agent transport.

Text
MCP, launched in late 2024, has already solidified its role as the tool-calling interface for AI agents. It defines how models interact with servers, enabling function discovery, invocation, and response interpretation over HTTP. The Linux Foundation reported 10,000 active public MCP servers and 164 million monthly Python SDK downloads by April 2026, confirming its dominance in this layer. A2A, donated to the Linux Foundation in June 2025, focuses on task coordination between agents, introducing mechanisms like Agent Cards and task lifecycle states. Enterprise teams have adopted A2A to address gaps left by MCP, according to Google’s documentation.

Text
ACP and ANP operate in separate domains. ACP, a lightweight message envelope format, simplifies agent-to-agent communication without the overhead of A2A’s task coordination. ANP, meanwhile, uses Decentralized Identifiers (DIDs) for agent identity and JSON-LD graphs for capability descriptions, enabling decentralized agent marketplaces. These protocols collectively form a stack: ANP for capability discovery, A2A for task coordination, MCP for tool calls, and ACP for lightweight messaging.

Text
The remaining challenge lies in the transport layer. All current protocols run over HTTP, a legacy of their origins in research, API, and enterprise environments where HTTP was the default. However, HTTP’s reliance on reachable servers creates bottlenecks for agents operating behind network address translation (NAT), which 88% of devices use. This necessitates relay infrastructure, adding latency, cost, and failure points.

Text
Solutions for NAT traversal and peer-to-peer (P2P) communication already exist. Technologies like UDP hole-punching with STUN, X25519 Diffie-Hellman encryption, and QUIC (RFC 9000) provide the building blocks for direct agent connections. However, these tools must be adapted for capability-based routing, where agents search for peers based on their functions rather than hostnames. Pilot Protocol’s IETF Internet-Draft and libp2p’s existing infrastructure are among the projects addressing this gap.

Text
Convergence in the transport layer is expected to lag behind application-layer protocols. While MCP and A2A are stabilizing, the transport layer will see 18–24 months of experimentation before consolidation. The IETF and W3C are likely to publish formal standards between 2027 and 2028, but open-source implementations may establish de facto norms earlier. Engineering leaders are advised to adopt application-layer protocols like MCP and A2A now, while treating transport-layer choices as temporary.

Text
Philip Stayetski, co-founder of Vulture Labs, emphasized the importance of separating application semantics from transport mechanisms. “Clean separation is cheap to implement now and expensive to retrofit later,” he said, citing lessons from the microservices era. Teams that design systems with this architecture will gain leverage as the transport layer stabilizes.

Text
The AI agent protocol landscape reflects a pattern seen in distributed computing history. Just as HTTP-native standards like REST replaced earlier, more complex protocols, the current proliferation may eventually narrow to a few foundational technologies. For now, the focus remains on layering solutions that address specific needs while preparing for broader interoperability.

Quoted text
“Clean separation is cheap to implement now and expensive to retrofit later.”
Source
Philip Stayetski, co-founder of Vulture Labs