Two of the world's largest AI powers spent the first half of 2026 answering the same question — how should autonomous agents find and talk to each other? — and reached for opposite instruments. That divergence, more than any single model release, is where the shape of the agent web is being decided.
In late June, China's State Administration for Market Regulation approved seven national standards for "Artificial Intelligence Agent Interconnection" — the country's first national standard set for how agents connect. The seven cover, in the regulator's own decomposition: general architecture, identity codes, identity management, agent description, agent discovery, agent interaction, and agent tool calling. Read that list as a stack, from the wire up to the tool call, and the interesting thing is not that it exists. It is the order.
The order is the argument#
Identity codes and identity management sit near the base of the stack — underneath description, underneath discovery. In plain terms: before an agent can be described, before it can be found, before another agent can call it, it must carry a recognized identity. The South China Morning Post described the outcome as "digital ID cards" for agents, folded into a unified identity-management regime whose stated goal is a "closed-loop" system for secure cross-domain interaction.
Now compare that to how the Western agent-interconnection layer actually got built. Model Context Protocol, which Anthropic published in late 2024, spread precisely because it asked for nothing at the door. You could stand up an MCP server in an afternoon and any client could call it. Discovery and tool calling came first; identity and authorization were someone else's problem, to be retrofitted later. And they are being retrofitted — loudly. MCP's own spec has been grinding through an OAuth overhaul, and a whole cottage industry of authority protocols has appeared this year, from x401 proving the human behind an action to web-bot-auth proving an agent is a known bot.
China standardized identity before capability. The West standardized capability and is arguing about identity after the fact.
That inversion is the story. It is easy to file this under "China regulates AI" and move on. That reading misses what is actually novel: China is standardizing the agent web as an identity-gated network, where legibility is a precondition of connection, not a feature added on top of it.
Market-first versus identity-first#
Neither approach is free. Identity-first interop buys governability — every agent is, by construction, attributable, and cross-domain traffic runs through a managed identity plane. What it spends is the thing that made MCP go viral: permissionless composability. The reason there are thousands of MCP servers is that nobody had to ask permission to publish one. A stack that requires a state-legible identity code before an agent can participate is, deliberately, a stack where fewer things connect by accident. That is a bug to a startup and a feature to a regulator, and both are being honest about which they are optimizing for.
The United States, for its part, has not issued anything comparable. In February 2026, NIST announced an "AI Agent Standards Initiative" aimed at interoperability and security — but it is a convening effort that seeks industry input, not a document vendors must build against. The de facto American standard remains whatever Anthropic, Google, and OpenAI ship and the market adopts. That is a governance philosophy too: let the interconnection layer emerge from competition and codify it afterward, if at all.
Why a domestic standard travels#
The obvious objection is that a Chinese national standard governs the Chinese market and stops at the border. True, and also beside the point. Standards are among the most durable strategic exports a state produces. Whoever defines the seven-layer decomposition that toolmakers, cloud providers, and agent frameworks build against is setting the default that the rest of the ecosystem either matches or has to translate around. China has said it intends to formulate dozens of AI-sector standards; agent interconnection is one brick in a much larger wall, and it is being laid early, while the category is still soft enough to shape.
The near-term effect inside China will be mundane and real: vendors will implement identity codes and discovery to conform, and a domestic agent-to-agent ecosystem will grow up already wearing ID badges. The long-term effect is the one to watch. For a decade the internet's foundational protocols were written in the open, by rough consensus, and everyone inherited them. The agent web is the first major networked layer being drafted in an era where two governance models — issued-and-legible versus emergent-and-permissionless — are competing to be the substrate. This month, one of them shipped seven documents. The other is still taking comments.
