On 26 June 2026 the State Administration for Market Regulation approved the release of seven national standards on “AI — Agent Interconnection”. As reported by Xinhua, with the rapid iteration of large-model technology, AI is accelerating from the perception-and-understanding stage into the generation-and-decision-making stage.
Why agent interconnection needs standards
An “agent” is an AI system that can autonomously perceive, plan and execute tasks. Unlike a model that only answers questions, an agent can call tools, access systems, chain multi-step operations and even collaborate with other agents.
When agents move from single to multiple, from isolated to collaborative, a foundational question arises: how do agents developed by different vendors and platforms recognise, communicate and coordinate with each other? Without common rules, each agent is self-contained and scale collaboration is impossible.
The “AI — Agent Interconnection” series provides a common language for this “interoperability” layer. The series uses a multi-part structure covering the overall architecture and establishing a technical baseline for interfaces, interaction and collaboration between agents.
The industry background
This release is not isolated but part of the accelerating construction of the AI standards system. Earlier, standards on AI terminal intelligence grading and large models were released; with agent-interconnection standards in place, agents move from “usable” to “collaborative” with a technical basis.
At the regulatory level, agents are also a focus. In earlier special actions, non-compliant agents were listed among the main inspection targets. Standards and regulation work in tandem — one provides technical rules, the other sets behavioural boundaries.
New compliance questions from agent scale-up
As agents move from single points to interconnection, compliance questions escalate.
First, responsibility. When multiple agents collaborate on a task and harm results, how is liability divided among developers, deployers and callers? This is more complex than for a single model service.
Second, permissions and access control. An agent that can call tools and access systems has “action capability”. Who authorises which resources it accesses, how boundaries are set and how overreach is blocked require dedicated design.
Third, data flow and personal information. Agent collaboration often involves data flowing across multiple entities and systems, requiring reassessment of personal information and important data processing arrangements.
Fourth, safety and controllability. The more autonomous an agent’s decision-making, the higher the requirements for monitoring, traceability and intervention, consistent with the risk-management philosophy of the AI safety governance framework.
Three pointers for companies
First, treat “interoperability” as a product-design premise. If the product plan includes multi-agent collaboration, interface and interaction design should align with the national standards as far as possible to reduce later integration costs.
Second, plan permission and liability mechanisms in parallel. As agents gain action capability, clarify authorisation scope, audit logs and liability boundaries to avoid being unable to attribute causes after an incident.
Third, watch the IP arrangements for agents. Agents calling tools, generating content and producing collaborative outputs raise ownership and licensing questions that must be arranged in advance. For support on agent compliance, data flow and liability arrangements, see our Data Compliance and Commercial Contracts services.
This article is based on public information on the ‘AI — Agent Interconnection’ national standards released by the State Administration for Market Regulation on 26 June 2026. It is for reference only and does not constitute legal advice. The specific content of the standards is subject to the official texts.