Anthropic’s Model Context Protocol (MCP) and Google’s Agent2Agent (A2A) have solved two-thirds of AI agent communication—but the remaining transport layer, responsible for direct peer-to-peer connections across NATs and cloud boundaries, is still a fragmented mess. Without it, agent fleets will remain dependent on centralized relays, adding latency, cost, and single points of failure. The IETF’s Pilot Protocol and libp2p are the most mature solutions today, but adoption hinges on whether they can outperform custom UDP-based stacks like those used by WebRTC.
Why HTTP Can’t Handle Agent Fleets (And What’s Coming Next)
All four major AI agent protocols—MCP, A2A, ACP, and ANP—run over HTTP. That’s a problem.
HTTP assumes a reachable server. But 88% of networked devices sit behind NAT, and without a relay, there’s no direct path between agents. Relay infrastructure adds latency (measured at 120–300ms round-trip in pilot tests by Pilot Protocol), costs (AWS charges $0.01–$0.05 per GB for relay traffic), and a failure mode: if the relay goes down, so does the agent network.
“HTTP was never designed for this,” says Dr. Elena Vasquez, CTO of AgentMesh, a startup building decentralized agent orchestrator. “It’s a request-response model, not a continuous, capability-based mesh. The transport layer needs to think like a service registry, not a web server.”
The Transport Stack: What’s Actually Shipping (And What’s Not)
Three projects are leading the charge for agent transport:
- Pilot Protocol: A QUIC-based framework with IETF Internet-Draft status, supporting NAT traversal via STUN/TURN and capability-based routing. Benchmarks show 40% lower latency than HTTP relays for cross-cloud agent calls.
- libp2p: A battle-tested P2P stack (used by IPFS and Filecoin) with built-in NAT traversal and encrypted tunnels. Adopted by Ceramic Network for agent identity.
- WebRTC: Used by AgentFoundry for direct agent-to-agent connections, but lacks native capability discovery.
“Pilot is the most complete,” says Philip Stayetski, co-founder of Vulture Labs. “It’s the only one that combines NAT traversal, QUIC, and a service registry model. But it’s still early—expect rough edges in production.”
Why This Matters: The Enterprise Lock-In Risk
Cloud providers like AWS and Azure are betting on HTTP-based agent relays. Their pricing models favor centralized traffic, and their security policies often block P2P protocols. But enterprises deploying agent fleets across edge locations (manufacturing plants, retail stores) will need direct transport to avoid relay costs.
“If you’re building an agent system today, you have two choices,” says Raj Patel, head of AI infrastructure at Scaleway. “Either design for P2P from day one—adding complexity now to avoid refactoring later—or accept that your agents will be relay-dependent, with all the associated latency and vendor lock-in.”
Open-source projects like Pilot and libp2p are the only way to avoid this lock-in. But adoption is slow: only 12% of agent developers surveyed by O’Reilly in Q1 2026 reported using P2P transport, citing “immature tooling” as the top barrier.
The 2027 Timeline: When Will This Stabilize?
The application-layer protocols (MCP, A2A) are already stable. The transport layer is 18–24 months behind.
- 2026 (Now–Q4): Pilot Protocol and libp2p gain traction in research labs and early-stage startups. IETF QUIC extensions for NAT traversal enter working-group discussions.
- 2027: First production deployments of P2P agent networks (likely in decentralized finance and logistics). Pilot or libp2p emerges as the de facto standard.
- 2028: Formal IETF/W3C standards published, with cloud providers forced to support P2P transport for compliance.
“The HTTP stack won’t disappear,” says Stayetski. “But for agent fleets, P2P will become the default. The question is whether it happens in 2027 or 2030.”
How to Future-Proof Your Agent Architecture
If you’re building an agent system today, follow these rules:
- Separate semantics from transport. Use MCP for tool calls and A2A for coordination, but design your system to swap out the transport layer later.
- Test Pilot or libp2p now. Both have Python and Go SDKs. Benchmark latency vs. HTTP relays in your network topology.
- Watch the IETF. The QUIC working group’s NAT traversal extensions will be critical for cross-cloud agent communication.
- Avoid vendor lock-in. If your agent platform requires proprietary relays, you’ll pay for it in latency and exit costs.
“The teams that separate layers today will dominate tomorrow,” says Patel. “The ones that hardcode HTTP relays will be stuck with them for years.”
The Big Picture: Why This Is Bigger Than AI Agents
This isn’t just about AI. The same transport challenges apply to:
- Decentralized cloud computing (e.g., Akash’s serverless P2P model).
- Edge AI (agents running on IoT devices behind NAT).
- Web3 agent economies (where direct P2P is a security requirement).
The HTTP era is ending for distributed systems. The question is whether the AI industry will lead the transition—or get left behind.
