MCP Co-creator Visits Japan, Unveils Long-Running Execution and Skills Vision

A conference focused on AI agents and the Model Context Protocol was held in Tokyo. The event was named AGNTCon+MCPCon Japan 2026. It took place on September 10 at an event hall in Shibuya, Tokyo. It was organized by the Linux Foundation. The keynote was delivered by David Soria Parra of Anthropic. He is one of the co-creators of MCP. His talk was titled MCP and the Era of Connectivity. According to reporting by jniino of Publickey, it was his first time participating in a conference in Japan.

Agent Conference Held in Shibuya

The conference was distinctive in its dedicated focus on agents and connectivity standards. Rather than product presentations by a single company, the agenda centered on developing open standards. The Linux Foundation’s role as organizer is seen as reflecting a call for neutral governance. The aim is to avoid dependence on any specific company and ensure interoperability. Developers and infrastructure specialists gathered at the venue, underscoring strong interest in agent infrastructure in Japan.

Parra emphasized the importance of open standards for connectivity. He argued that standards are indispensable as part of the AI stack. Improvements in standalone model performance alone will not drive integration into real-world operations. A safe pathway for連携 with existing business systems is needed. MCP, he said, is positioned to provide that pathway.

His explanation began with handling long-term work. The focus was on an area called Long-Horizon Tasks. This refers to complex work spanning dozens to hundreds of steps. Such tasks can take from tens of minutes to several days. Charts illustrating improved model performance were shown at the venue.

Evolution of Models Handling Long-Running

Processes

The scale of work that models can handle continues to expand. The stage where months of human labor can be processed in hours has come into view. Parra said he expects such cases to emerge within months. Development practices for agent-related products are also changing. Designing for long-running processes has become realistic.

Looking back at 2024, tool calling was a new concept. That coincided with the period when MCP was created. Models could not reliably handle long tasks. Tool-calling capabilities were also immature. They were not ready for use in production environments.

The turning point came in 2025. Parra described that year as the year of the coding agent. It was the first practical agent to see real-world adoption. The reason lies in its compatibility with automation systems. Behavior could be guided using unit tests and translation systems. A loop for detecting errors and making corrections could be built in. Practical value could be secured even with only a few hours of continuous operation.

At present, model capabilities have improved significantly. There are more cases where models can reach correct results on their own. Parra said we have entered an era in which long-running work can be delegated. He expressed the view that 2026 will be the year when use in production environments expands. Applications will expand beyond coding. Knowledge work, scientific discovery, mathematics, and other fields will be covered.

2026 will be the year when agentic AI systems are used in real production environments for the first time, expanding from coding into diverse fields including general knowledge work, scientific discovery, and mathematics

The above is a summary of the outlook presented in the talk. It suggests a shift from the stage of coding assistance to the core of business operations.

Connectivity Shaping Knowledge Work

Requirements differ between coding and knowledge workers. In coding, verification infrastructure served as a guide. For knowledge work, by contrast, connectivity is critical. Parra clearly distinguished that difference. Professions such as financial advice and accounting were cited as examples. Integration with Microsoft 365 and Google Workspace is essential. Back-and-forth with business systems and issue-tracking systems occurs daily.

As models’ composite capabilities grow, the design of connections becomes more important. Parra said the real key lies in how to connect. This trend is visible across the ecosystem. Dependence on standard specifications for tool integration is increasing. Maintaining individual integrations cannot ensure scalability. Connection through common standards reduces cost and maintenance burdens.

From the perspectives of safety and manageability, connection standards are also meaningful. Handling of permissions and audit trails can be standardized. It becomes easier for companies to decide whether to adopt them. Such infrastructure development is a prerequisite for production use. It is not unrelated to the safety debate highlighted in Anthropic Researcher Warns AI Human Extinction Probability Exceeds 10%, as improved capabilities and expanded connectivity demand mechanisms for control.

MCP Surpasses 1 Billion Cumulative Calls

The expansion in usage scale was presented in concrete figures. Tool calls via MCP on Claude have cumulatively surpassed approximately 1 billion. Monthly downloads of SDKs for MCP exceeded 500 million. Usage has surged rapidly, from Anthropic to its competitor OpenAI. A small project started about two years ago has become industry infrastructure.

This rapid growth can be seen as a success story for standardization. It is the result of developers choosing a common interface over individual implementations. For tool providers, distribution channels have also been simplified. There is an advantage in being used by multiple applications with a single integration. For users, switching costs are lowered.

On the other hand, concentrated dependence increases the responsibility on the maintenance side. The impact of specification changes widens. Compatibility and migration guidance are required. The importance of a neutral governance structure is growing. This also overlaps with the debate over release management seen in Anthropic to Release Ultra-Dangerous AI Claude Mythos on a Limited Basis, as more powerful features require phased release and verification.

Specification Evolution Toward Stateless Operation

MCP has expanded its features over about two years since its launch. Remote connection capabilities were added and became widely used. Last year, authorization features were added. Features such as MCP Apps were also developed. And this year, a shift to fully stateless connections was implemented. It was described as the largest and most important change in the past 18 months.

This change was advanced through cross-industry deliberation. Microsoft, Google, Anthropic, OpenAI, and others participated. A challenge was dialogue patterns that are difficult to handle with HTTP alone. Support was needed for requests requiring multiple round trips. It was necessary to support dialogue patterns specific to agentic systems.

Stateless connections offer major operational advantages. Scalability and recovery from failures become easier. Design of load balancing and retries is simplified. Parra said MCP is currently in good shape. However, preparation is needed for adoption over the next one to three years. It is preparation for an era when it becomes the standard means for knowledge work.

MCP Tasks Concept for Long-Running Execution

Focus areas for the next 12 months were also presented. The first is developing agentic information exchange. The aim is to make MCP the standard procedure for information exchange and orchestration. As part of this, the concept of MCP Tasks was introduced. It is a mechanism for invoking an MCP server as an agent. It allows delegation of processes lasting minutes or hours. Long-running processes spanning weeks or months are also envisioned. It takes the form of returning results upon completion.

Long-running execution differs in design from conventional request-response. Handling progress tracking, interruption, and resumption becomes a challenge. Management of permission expiration and costs is also needed. The Tasks concept is seen as addressing such operational requirements. In work other than coding, waiting times tend to be longer. This is because they include waiting for approvals and materials. Asynchronous execution management becomes a condition for real-world adoption.

This direction also affects product design. Applications will no longer be able to assume immediate responses. Design of status displays and notifications becomes important. On the infrastructure side, persistence of execution and auditing are required. Reliability of long-running processes will determine adoption.

New Extension Aimed at Skill Distribution

The second pillar is enriching semantic content. It will increase the semantic descriptions that can be handled within MCP servers. It will allow issues and intentions to be expressed more richly. Parra said this is the feature he personally looks forward to most. It is an extension called Skills over MCP. It provides AI agent skills via MCP. It aims for organic support across the Claude product family and the industry as a whole.

The aim is to bundle and deliver tools and skills together. They can be provided together on an MCP server. Distribution without reliance on complex marketplace mechanisms becomes possible. It also opens the way to unified provision of in-house skills. Expertise within companies can be organized as skills. It has the effect of making reuse across departments easier.

Distribution of skills also involves quality management challenges. Indication of origin, version, and permissions is essential. It is necessary to prevent the inclusion of malicious or outdated skills. If provenance information can be handled under a common standard, review becomes easier. This also relates to the tension between openness and regulation discussed in Anthropic Mythos Export Controls Risk Repeating PGP Mistakes, as ease of distribution and control must both be achieved.

Editorial Opinion

On short-term impact, over the next three to six months we expect to see an increase in prototyping of long-running execution infrastructure. Following the shift to stateless connections will be an immediate task for developers. Work to update SDKs and receiving sides will intensify. Bundled delivery of skills will also begin trial introduction. We assess that inventorying internal skills and organizing permissions will come first.

From a long-term perspective, the focus will be on rewriting knowledge-work infrastructure. We expect connection standards to become a prerequisite for business processes within one to three years. Migration from individual integrations to common interfaces will proceed. Standardization of auditing and authorization will speed up adoption reviews. Reuse of skills will determine organizational productivity. Stable operation of standards will become a competitive condition, in our assessment.

We note questions from the editorial team. Is the division of responsibility for long-running execution clear. Who bears the cost of re-execution upon failure. Will provenance labeling for skills function adequately. Can the neutrality of open standards be maintained. Is there a concern that a specific company’s implementation will become the de facto standard. Verification of these issues will be required in future coverage.

References

• “AnthropicのMCP共同作者が来日、基調講演で語るこれからのMCP注力領域。AGNTCon＋MCPCon Japan 2026”, by jniino — Publickey, 2026-09-14T15:55:18.000Z (ARR)
• Source URL: https://www.publickey1.jp/blog/26/anthropicmcpmcpagntconmcpcon_japan_2026.html