AI Agent Protocols Evolve, Tackling Tool Calling and Task Coordination While Addressing Transport Challenges

The evolution of distributed computing has historically involved an initial proliferation of protocols followed by a period of consolidation. Examples include the competition between CORBA, DCOM, RMI, and SOAP before REST became dominant for enterprise integration, and the fragmentation in real-time messaging before MQTT and WebSockets established their niches. The current AI agent ecosystem is mirroring this pattern, with a burst of new protocols emerging.

Over the past eighteen months, four significant protocols have been published. Anthropic's Model Context Protocol (MCP), released in late 2024, functions as a tool-calling interface, defining how models interact with server functions. By April 2026, the Linux Foundation reported over 10,000 active public MCP servers and 164 million monthly Python SDK downloads, indicating its established position in the tool-calling layer. IBM Research introduced the Agent Communication Protocol (ACP) in March 2025, a lightweight, stateless message envelope format for agent-to-agent exchanges. Google's Agent2Agent (A2A), published in April 2025 and later donated to the Linux Foundation in June 2025, serves as a task coordination interface. It enables agents to delegate tasks, featuring Agent Cards for capability advertisements, task lifecycle states, and three interaction modes: synchronous, streaming, and asynchronous. An independent working group developed the Agent Network Protocol (ANP), which focuses on discovery and identity using Decentralized Identifiers (DIDs) and JSON-LD graphs.

These protocols address different layers of the AI agent communication stack, complementing rather than directly competing. ANP or simpler registries handle capability discovery, A2A manages task coordination, MCP facilitates tool calls, and ACP provides lightweight messaging where full task lifecycle management is not required. However, a significant challenge persists at the transport layer, as these protocols primarily operate over HTTP.

The reliance on HTTP presents limitations for agents requiring direct peer-to-peer communication, especially for the 88% of networked devices behind Network Address Translation (NAT). HTTP assumes a reachable server, necessitating relay infrastructure for NAT traversal, which introduces latency, cost, and potential failure points. This transport problem, distinct from the application-layer semantics addressed by MCP, A2A, ACP, and ANP, is a session-layer issue.

Existing technologies can address this problem, including UDP hole-punching with STUN for NAT traversal (effective for approximately 70% of network topologies), X25519 Diffie-Hellman and AES-256-GCM for authenticated encryption, and QUIC (RFC 9000) or custom sliding-window protocols over UDP for reliable delivery without TCP's head-of-line blocking. Projects like Pilot Protocol, libp2p, and the IETF's QUIC working group are developing solutions, with a focus on capability-based routing where agents find peers based on their functions rather than hostnames.

The HTTP-based application protocols like MCP and A2A are nearing stable versions, with the next 12 months expected to bring production hardening and security improvements. The transport layer is projected to converge 18 to 24 months later, following a period of diverse implementations and empirical data accumulation. Formal standardization from organizations like the IETF and W3C is anticipated between 2027 and 2028. For engineering leaders, current architectural decisions should include adopting stable application-layer protocols while designing agent systems with a clear separation between application semantics and the evolving transport layer.

According to VentureBeat, Philip Stayetski is a co-founder of Vulture Labs.