Anthropic released the Model Context Protocol in late 2024 as an open standard for connecting AI agents to external tools and data sources. Google followed in April 2025 with the Agent-to-Agent protocol, which standardizes how agents delegate tasks to one another. Both were designed to be adopted freely, with no licensing requirements, no certification programs, and no governance obligations attached. By mid-2026, the two protocols had attracted adoption from every major cloud provider and AI lab. In September, three companies on three continents each launched enterprise agent platforms built on those protocols within two weeks of one another. The infrastructure beneath the shared protocol layer had almost nothing in common.
Universal adoption
OpenAI opened its Agents API in public beta on September 10, giving all developers the managed orchestration harness that powers Codex and ChatGPT for Work. The API provides durable sessions, tool orchestration, subagent coordination, and a choice between managed and self-hosted execution environments. Huawei followed at CONNECT in Shanghai on September 18, launching AgentArts and its open-source counterpart openJiuwen with over 5,000 general-purpose MCP assets and 1,000 industry-specific ones. Alibaba unveiled AgentCore at the Apsara Conference on September 22, an enterprise platform for building, running, and governing agents throughout their lifecycle, with centralized security and audit controls built into the governance layer.
All three adopted MCP as their standard for connecting agents to external tools and data sources. Huawei’s openJiuwen ships C++ SDKs for both MCP and A2A, implementing the Western-originated interoperability stack on Chinese silicon. Alibaba’s AgentCore manages MCP servers as first-class objects within its governance layer, alongside models and agent skills. The protocol vocabulary is shared across all three ecosystems.
Stack divergence
Huawei’s agent infrastructure runs entirely on its own Ascend processors, organized into cluster architectures linked by a proprietary interconnect. ThinkPro, a new runtime announced at the same event, treats agent processes as fundamental units within Huawei’s own Linux distribution, embedding agentic execution at the operating system level. The Ascend developer community receives shared access to 10,000-NPU-scale compute clusters for development and experimentation. Huawei reported that the openJiuwen community has surpassed 50,000 GitHub stars and 3.29 million downloads. No component of this stack originates from or depends on a Western vendor.
Alibaba’s AgentCore operates within a layered cloud architecture that separates model training, agent deployment, and contextual data management. A component that Alibaba calls the Context Engine connects enterprise documents, business systems, and chat records to agents in real time, claiming a 67% reduction in token usage for knowledge-intensive applications such as customer service and analytics. The platform already serves more than 100 enterprise customers in China.
OpenAI’s Agents API defaults to GPT-6 Astra on managed sandboxes partnered with Cloudflare and Oracle, with data residency restricted to the United States. The three platforms run on different processors, train different models, deploy in different markets, and operate under different data sovereignty rules. MCP is the only surface on which they converge.
Free forever
MCP reached 97 million monthly SDK downloads by early 2026. Adoption required no licensing cost, no certification, and no ongoing relationship with Anthropic. Google took the same approach with A2A, donating the specification to the Linux Foundation in June 2025. By April 2026, the protocol had reached v1.0 with backing from over 150 organizations including AWS, Microsoft, Salesforce, SAP, and ServiceNow. The frictionless adoption that made both protocols dominant also ensured that their creators retained no governance authority over how adopters used them.
Huawei deploys those 5,000 MCP assets across government, utilities, and healthcare on Ascend hardware that operates entirely outside US export controls. Alibaba’s AgentCore manages MCP servers for more than 100 enterprises across Chinese industry. The protocols’ creators have no presence in these sectors, no contractual relationship with these adopters, and no mechanism to influence these deployments. The protocols were designed to require none of those things. Maximizing adoption by minimizing control produced the global reach that placed the protocols beyond their inventors’ governance.
Permission not required
HTTP did not remain under CERN’s direction, and TCP/IP did not remain under DARPA’s. Both became global infrastructure through unrestricted adoption, and their creators’ influence diminished in direct proportion to their reach. MCP and A2A have followed the same trajectory on a compressed timeline, reaching worldwide deployment within two years of their initial release. The agent stack now develops simultaneously in Shenzhen, Hangzhou, and San Francisco, on hardware and within markets that the protocols’ creators neither supply nor access. Those protocols will continue to shape the agent infrastructure layer precisely because they were engineered to be adopted without permission or obligation.