From late 2024 into 2025, three agent communication standards landed nearly simultaneously: Anthropic's MCP, Google's A2A, and the community-built ANP. The pattern echoes the TCP/IP battles of the 1990s — once a standard is set, every latecomer has to follow.

What this is

In plain terms, these three protocols solve different layers of the "AI calling AI" problem.

  • MCP (Model Context Protocol): Proposed by Anthropic, it defines how AI calls external tools — querying databases, hitting GitHub, reading files. To add a new capability to an AI assistant, you don't rewrite code; you plug in the protocol.
  • A2A (Agent-to-Agent Protocol): Proposed by Google, it defines how AIs collaborate with each other. A "researcher" AI hands material to a "writer" AI, which is then polished by an "editor" AI.
  • ANP (Agent Network Protocol): Community-led, targeting larger-scale agent networks — essentially an "internet" between agents.

Industry view

Supporters argue that protocolization is a prerequisite for AI to scale. Anthropic and Google are both contesting the "standard-setter" position — whoever's spec gets adopted by more developers controls the underlying rules for AI product interoperability.

But the risks are clear. First, fragmentation: the three protocols are currently incompatible, and enterprises that back the wrong camp pay migration costs. Second, fears of enclosure: MCP and A2A are both steered by large companies, and the community worries they will become new "walled gardens." Third, unproven commercial value: the protocols themselves don't make money directly. The ultimate winners may be the upper-layer applications, not the standard-setters — recall the browser wars of the past: real money was made by search engines and e-commerce, not by the browsers themselves.

Impact on regular people

For enterprise IT: When procuring AI assistants in the future, "which protocols does it support" will become an evaluation criterion, just like today asking "which databases does it support."

For individual careers: Products built around multi-AI workflows are coming online. The division of labor for research, writing, analysis, and similar tasks will be redesigned — but this shift will take 1–2 years to reach ordinary positions.

For consumer markets: No visible change in the short term. It takes 2–3 years for underlying protocols to surface in consumer products, at which point assistants like Siri and Xiao Ai will see noticeably expanded capability boundaries.