Anthropic's David Soria Parra: 2026 is agentic AI's first production year

MCP was created in 2024, when tool calling from an AI model was still a very new concept, Parra said. At the time, models could not handle long tasks and tool calling was not sufficient for production use. He described 2025 as the year of coding agents, and explained that coding agents were the first practical agents in production because automated systems such as unit tests and compilers could steer and correct a model's behavior. Now that models can handle longer tasks on their own, Parra expects the applications to widen beyond coding.

He drew a line between software developer tools and tools for general knowledge workers. For knowledge workers, the key is connectivity: financial advisors and accountants, for example, need to work across systems such as Microsoft 365, Google Workspace, internal business systems and task management tools. Parra said this is why MCP and similar standards are spreading, and he pointed to the volume of tool calls on Claude and monthly MCP SDK downloads as evidence that the whole industry now depends on what was a small project two years ago.

The change Parra highlighted as the largest in the past 18 months is making MCP fully stateless, an effort he said involved leaders from Microsoft, Google, Anthropic and OpenAI working together. Looking ahead, the outcome appears to hinge on the next 12 months of work: whether agentic messaging, richer semantics and authorization and identity mature enough to support his expectation that agentic systems become a standard means of knowledge work within one to three years. He closed by inviting developers to build and share MCP servers, bring Japan's market issues to conferences, and join open-source communities to shape the standards.