HTTP agent protocols solved semantics, but NAT makes the transport layer the real bottleneck

AI agent protocols have a plot twist: the “what to say” parts are getting standardized fast, but the “how to connect” part is still stuck behind NAT. VentureBeat’s breakdown is blunt about the mismatch. MCP, A2A, ACP, and ANP all run over HTTP, and HTTP quietly assumes you can reach a server. In the real world, that assumption breaks. The result is that agent fleets that need direct peer-to-peer routing across cloud boundaries, home networks, and edge deployments end up forced through relay infrastructure, adding latency, cost, and an extra failure mode.

So what actually solves the transport problem for AI agents? The answer is not “another application protocol.” It’s a session-layer and tunneling problem, and the source points to the same primitives used elsewhere on the internet: UDP hole-punching with STUN, authenticated encryption using X25519 Diffie-Hellman and AES-256-GCM, and reliable delivery over UDP via QUIC (RFC 9000) or custom sliding-window protocols. Those building blocks exist. What’s missing is capability-based routing for agents, where discovery is based on what peers can do, not just where they live.

To understand why this is happening now, zoom out to how distributed computing usually evolves. VentureBeat compares the current AI agent ecosystem to past waves of protocol proliferation followed by consolidation. In the late 1990s, enterprise integration bounced between CORBA, DCOM, Java RMI, and early SOAP before REST and HTTP won through simplicity and HTTP-native tooling. Real-time messaging fragmented too, with XMPP and IRC and a stack of proprietary protocols, until MQTT and WebSockets carved out niches. That pattern matters for executives because it sets expectations: the market often standardizes semantics first, then converges on transport only after enough deployments reveal what actually works.

This time, the published protocols are coming quickly. VentureBeat lists four significant agent-focused standards published in the past eighteen months: Model context protocol (MCP) from Anthropic in late 2024, agent communication protocol (ACP) from IBM Research in March 2025, Agent2Agent (A2A) from Google in April 2025, and agent network protocol (ANP) from an independent working group. It also notes that the W3C AI Agent Protocol Community Group has opened a standards track, while the IETF is receiving Internet-Drafts on agent transport. The momentum is visible in the code and the conferences too, with workshops on interoperability and a steady stream of GitHub repos claiming they solve agent communication.

But not all protocols compete for the same slot. The source is careful here, and the clarity is the real payoff for busy leaders: MCP is a tool-calling interface, a typed remote procedure call contract between a model client and a tool server, running over HTTP. A Linux Foundation confirmation cited by VentureBeat says there are more than 10,000 active public MCP servers and 164 million monthly Python SDK downloads by April 2026, and it adds that MCP has already won the tool-calling layer with “standardization work effectively done.”

A2A is different. It’s task coordination for multi-agent systems, defining how agents delegate tasks using Agent Cards for capability advertisements, task lifecycle states, and three interaction modes: synchronous, streaming, and asynchronous. VentureBeat also states Google donated A2A to the Linux Foundation in June 2025, and that enterprise AI teams have adopted it broadly because it fills the coordination gap that MCP leaves open.

ACP is positioned as a message envelope for agent-to-agent message exchange without A2A’s full coordination semantics, suitable when lightweight message passing is enough. ANP is discovery and identity, using Decentralized Identifiers (DIDs) for agent identity and JSON-LD graphs for capability descriptions. The emerging stack therefore looks layered: capability discovery via ANP (or simpler registries), task coordination via A2A, tool calls via MCP, and lightweight messaging via ACP when you do not need full task lifecycle management. The layers complement rather than compete. That makes the strategy simpler. Your architecture should treat these as application-layer semantics that can evolve on a timeline different from transport.

Which brings us back to the transport layer, where VentureBeat says the ecosystem is 18 to 24 months behind. The core issue is not “security” or “message format.” It’s routing and connectivity. NAT means that without a relay there is no reachable server, and the source claims 88% of networked devices sit behind NAT. HTTP-based protocols do not solve session setup or direct connection establishment, which is described as a Layer 5 problem in the OSI model. That gap forces relay infrastructure into the data path, pushing costs and latency up and creating a new dependency.

For engineering teams, the next 12 months are more about hardening than reinventing. VentureBeat expects production hardening and security improvements, along with “stateless MCP servers for horizontal scaling” and better A2A federation. It also expects a period of implementation diversity for transport because teams will experiment with different peer-to-peer agent networking approaches before consolidation around a small number of implementations. Standardization will likely land in a 2027 to 2028 window via the IETF and W3C tracks, with one or two open-source implementations potentially gaining de facto status before formal specs arrive.

The second-order implication for leadership is strategic separation. VentureBeat’s practical recommendation is not a new protocol. It is a design principle: clean separation between application semantics (MCP, A2A) and transport (whatever sits below). That separation is cheap to implement now and expensive to retrofit later. If you build tightly coupled to HTTP transport, you risk paying a major rewrite when transport stabilizes. If you architect in layers, you can adopt stable semantics now while you watch the transport landscape evolve from experimentation to consolidation.

Net: MCP solved tool calling, A2A solved coordination, and ACP and ANP complete important pieces. But the real competitive advantage may belong to teams that treat transport as a replaceable component and start planning for capability-based peer-to-peer connectivity now.