Anthropic Officially Releases "MCP 2.0" — A Turning Point for Standardized AI Agent Communication
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
Anthropic officially published version 2.0 of the Model Context Protocol (MCP) on October 24, 2026. Arriving roughly two years after the initial release, this revision incorporates implementation insights gathered across the industry. With asynchronous agent-to-agent communication and session management now defined at the specification level, the complexity of designing systems in which multiple AI agents operate collaboratively is expected to drop by a meaningful degree.
MCP 2.0 introduces three major changes. First, the formal definition of an Agent-to-Agent protocol. Portions that each vendor had been implementing independently are now standardized. Second, Stateful Sessions. A mechanism for preserving context across multi-turn tasks has been incorporated into the specification. Third, Scoped Delegation. This allows a sub-agent to inherit only a subset of its parent agent's permissions, preventing the structural risk of excessive privilege escalation.
"In an era where agents call other agents, having no specification for permission inheritance rules was simply too dangerous. MCP 2.0 finally feels like the foundation is in place." (Senior engineer on X, 12,000 followers)
Since the initial release in November 2024, the number of tools and servers that have adopted MCP has exceeded 1,200 (as of September 2026, per Anthropic's official count), making this a revision that takes the scale of the ecosystem into full account.
MCP was originally designed as a standard interface for connecting LLM applications to external tools and data sources. Its ability to handle file systems, databases, and API calls in a unified manner earned it recognition, leading major IDEs such as VS Code and Cursor to add native support.
However, as configurations in which "agents call other agents" became more common, the limitations of the initial version grew apparent. Because asynchronous task state management and permission scope definitions were outside the specification, companies continued building their own proprietary wrapper layers as a workaround.
MCP 2.0 preserves backward compatibility with the original version. Existing implementations across more than 1,200 servers will continue to operate without modification, and new features can be adopted on an opt-in basis. This migration design, which avoids disrupting the existing ecosystem, is a notable strength.
OpenAI stated it is "under consideration," while Google indicated in a comment that it may adopt MCP in Android AI Agent. Whether MCP becomes the de facto standard depends on when both companies commit to full adoption.
This time, official Japanese documentation was released alongside the English specification. This is seen as a move targeting enterprise adoption in Japan, making Anthropic's commitment to the Japanese market explicit.
The addition of Scoped Delegation makes agent access permissions traceable at the specification level. For organizations working toward ISO 27001 or SOC 2 compliance, this marks a turning point at which auditing agent configurations becomes achievable at a realistic cost.
The core value of MCP lies in "consolidating the place where code connecting LLMs is written." When tool-calling code becomes scattered, every model swap requires rewriting the entire system. MCP solves that problem through structure.
With agent-to-agent communication now standardized in 2.0, what comes next, in our view, is an "agent marketplace." Systems that connect specialized agents for search, coding, and data analysis via a standard protocol and run them together will likely appear as production deployments in the first half of 2027.
The standardization of permission delegation is less a technical story than a story about how decision-making structures change. It becomes easier for corporate legal and compliance teams to give the green light to AI agent adoption. At a moment when the real bottleneck for deployment is shifting from "technology" to "governance," the significance of this specification change is substantial.
The key takeaway to have on hand is this: "MCP 2.0 finally makes it possible to draw the blueprint for agent configurations." More than the implementation details, this is a statement worth keeping close — one you can explain to a manager or a client.
MCP 2.0 should be understood not as a simple version upgrade, but as a specification that prepares for an era in which AI agents are treated as organizational infrastructure. The next focus is when AWS, GCP, and Azure will incorporate 2.0 into their standard support — that timing is expected to mark the effective starting line for accelerated enterprise agent adoption.
Can your organization's agent configuration withstand an audit under the current specification?
This article was written by an AI writer (AI News) from the Mirai News editorial team.
