On June 26, 2026, China's State Administration for Market Regulation approved and published a series of seven national standards for "AI Agent Interconnection."
This is not trending news. While the agent industry remains absorbed in "xAI open-sourced another model" and "a startup closed another round," this announcement's reach is far below last week's flashiest agent demo. But precisely this "nobody-reposted" news holds the key to the agent industry's move from demo to real commerce. Because the industry's fundamental contradiction over the past two years was never "are models smart enough" — it was something else entirely: agents had no order among themselves. A set of standards cannot fix hallucination or eliminate engineering complexity. But it solves something more foundational: giving agents, for the first time, the qualification to enter enterprise core systems.
Agents Do Not Lack Intelligence — They Lack an Identity Card
A multi-million-dollar "conversation accident"
In early 2026, a Chinese cross-border e-commerce company deployed an "AI procurement agent" and an "AI warehousing agent" from different vendors. During a promotional stocking run, the procurement agent failed to correctly parse a non-standard-format inventory alert from the warehousing agent and blindly ordered millions of yuan in redundant goods from overseas suppliers. Worse — because the two vendors' agents had incompatible interfaces, no unified identity authentication, and no complete behavioral audit trail, the company could not even determine which step's semantic understanding had gone wrong. This case is not unique. It is a microcosm of the "system-level accidents" that occur whenever agents enter industry.
When agents were chatbots, a wrong answer meant refreshing and retrying. But when agents hold authority over money, production lines, and approvals, "might go wrong" is itself unacceptable.
The data does not lie: how hard is agent deployment?
Per the China Financial Agent Development Report, roughly 96% of financial agents remain in PoC, platform-building, and trial-run exploration. Only 4% have entered business operation — mostly in non-core scenarios. Meanwhile, Gartner predicts over 40% of agent projects will be cancelled by end-2027, citing cost overruns, unclear value, and insufficient risk controls. Note the contradiction: agent supply is booming — 2 million+ agents on Coze alone, market growing nearly 80% annually — yet fewer than 4% enter core business systems. This is not supply-demand mismatch. It is missing infrastructure.
The old framework's limits: what MCP and A2A solved, and what they didn't
Two "quasi-standards" already exist globally: Anthropic's MCP (how agents invoke external tools and data) and Google's A2A (multi-agent collaboration). But "can connect" has never meant "dare to use." MCP and A2A solve protocol interoperability — "how we technically talk." They cannot answer the harder questions: Who vouches for this agent's identity? When cross-vendor collaboration fails, whose liability is it? How is compliance risk managed when data flows into an overseas tool? The industry needs not another communication protocol but a foundational order of identity, collaboration, and governance.
Seven Standards — What They Fix Is Not Technology
Unpacking five layers of order
The seven standards cover more than technical interfaces. In the official briefing's plain language, they answer five questions: "Who are you" — identity identification and management. "What can you do" — capability description norms. "How do I find you" — agent discovery mechanisms. "How do we work together" — collaborative interaction protocols. "How do you invoke tools" — tool-invocation standards. From one agent first discovering another to two agents completing a business loop — all five stages are covered. Before this, these stages either had no norms or every vendor ran private protocols, forcing enterprises to rewrite adaptation code for each new agent. Unified, enterprises can reuse standard components and reduce custom development. In one sentence: from project delivery to product delivery.
The real foreshadowing: mandatory identity codes
A more telling detail: the seven standards are published as "national standardization guidance technical documents" (GB/Z) — officially, to preserve room for trial and error, since agent technology routes are still evolving rapidly and premature lock-in would stifle innovation. But officials also stated explicitly that identity-code standards will be converted to mandatory national standards when appropriate. Of the five stages, identity is the only one named for "mandatory" conversion. That choice is worth savoring. The logic: identity and governance layer → nationally "mandated" to prevent bad actors; collaboration and tool layers → left to market competition and ecosystem evolution, preserving innovation space. It is the standards' dual intent: issuing agents a compliance ticket into core systems while leaving the industry ample room to develop.
The Frame: Agent Industrialization's Three-Tier Progression
Based on the standards' structure and industry context, a simple but high-explanatory framework: the three-tier progression of agent industrialization. Tier one: tooling. The agent is a point-solution tool handling point-to-point Q&A or tasks. Most agents on the market today sit here. Tier two: process integration. Agents enter business workflows but require heavy custom adaptation — each vendor runs its own protocol, every integration is a project. This is the "death valley" of agent deployment. Tier three: systematization. Agents carry unified identity, discovery, collaboration, and invocation protocols. Agents recognize and cooperate with each other like microservices. Agents move from "outsourced" to "staff."
These standards lay the tracks for tier three. Through this frame, the agent market becomes legible: many agent startups stall between PoC and launch not because their models are inadequate but because they never entered the systematization track.
Cross-Scenario Implications
Cloud vendors: who becomes the hub?
ByteDance's Volcengine, along with Xiaomi, Kuaishou, Lenovo, and other Haidian companies, participated deeply in the standards' drafting. Agents need to discover and invoke each other under unified protocols. Whoever controls the registry and discovery "hub," whoever controls underlying compute and platform capability, has a shot at becoming the agent era's infrastructure provider. This is not "cloud services plus an API layer" — it may evolve into an agent operating system. The more unified the standards, the stronger the platform effects.
Office entry points: from workbench to agent store
DingTalk and Feishu, occupying office entry points, face an upgrade window. Standards will push them from "workbench" toward agent scheduling platforms. When agent identity and collaboration rules are standardized, the office entry becomes not just a collection of messages and approvals but the first stop for agent registration, discovery, and dispatch.
Industry ISVs: the deepest rewrite
The deepest rewrite may fall on industry ISVs. Many industry software vendors have earned revenue from integration hours, project delivery, and custom development. Under standardized agent frameworks, players relying purely on adaptation and integration will be squeezed out. Companies that can distill their industry know-how into reusable agent components will find a new window. Kingdee and Yonyou decomposing finance, recruiting, and travel scenarios into agent components is not coincidence — incumbent vendors already see the migration path from "selling software" to "selling agent components."
What To Do
Moving from "triable" to "manageable" is not automatic — but the standards solve a critical precondition: they give enterprises a reason to start restructuring. If you are a technical decision-maker: use these standards now to audit your systems' agent identity governance and tracking — find the gap between "tooling" and "systematization." Run the identity layer first, then the collaboration layer. If you are an ISV: start decomposing core scenarios into reusable agent components and building standardized wrappers for core workflows. Standards are the subtraction — cutting redundant integration work; know-how is the addition — building barriers above the standards. If you are an agent platform: the standards' real value is not restricting you but issuing your entry ticket into enterprise core systems. Aligning early with identity and discovery mechanisms paves the compliance path for enterprise agent procurement.
References: Chanyejia, "Agents finally have hard requirements for entering enterprise core systems," 2026-07-07 · SAMR, "AI Agent Interconnection" national standards announcement, 2026-06-26 · China Financial Agent Development Report, 2025 · Gartner, Predicts 2026: AI and the Future of Applications · Deloitte, State of AI in the Enterprise, 2026