David Soria Parra on the Future of AI Agents

David Soria Parra, a Member of Technical Staff at Anthropic, recently took the stage at AI Engineer Europe to discuss the evolving capabilities and future of AI agents. His presentation, titled "The Future of MCP," explored the progression from early AI demos to sophisticated agents capable of complex tasks and programmatic interaction.

StartupHub data

Companies working on this

Profiles of the companies named in this story, with funding and a one-liner from our database.

Anthropic

Private / $100B+ est

Anthropic is an AI safety and research company building reliable, interpretable, and steerable AI systems, best known for the Claude family of models.

The Evolution of AI Agents

Parra began by highlighting the rapid development in the field of AI agents. He recalled that just a year prior, the concept was largely confined to specification documents and local tools. However, the past 18 months have seen a significant acceleration, with agents now capable of sophisticated actions like shipping their own interfaces through protocols.

The rapid advancement is evidenced by the sheer volume of development. Parra noted that in the last twelve months, the MCP (Model Context Protocol) has seen tremendous growth, with over 110 million SDK downloads per month. This indicates a strong developer interest and adoption of the underlying framework.

Key Milestones in Agent Development

Parra presented a timeline of MCP development, marking key milestones:

• November 2024: Open-sourcing of the core MCP.
• March 2025: Introduction of remote server capabilities.
• June 2025: Implementation of authorization mechanisms.
• September 2025: Development of elicitation primitives.
• December 2025: Integration of task management features.
• Q1 2026: Launch of MCP applications.

This progression signifies a shift from basic functionality to a more robust and integrated system, supporting hundreds of clients and thousands of servers with a single specification.

The Role of Connectivity and Protocols

A central theme of Parra's talk was the importance of connectivity and the right tools for building effective AI agents. He emphasized that agents need to interact seamlessly with various systems and data sources. This is achieved through a "connectivity stack" comprising:

• Skills: Domain knowledge captured as reusable instructions, allowing agents to perform specific capabilities. These skills are largely reusable, though some platform dependencies may exist.
• MCP: The core integration protocol that provides semantics, governance, and cross-boundary reach. It acts as the connective tissue between different components.
• CLI / Computer Use: General access to existing systems, enabling discoverable, composable, and use-style interactions.

Parra stated, "These compose. Agents in 2026 use all of them." This highlights the integrated nature of future AI agents, which will draw upon all these elements to perform their tasks.

The 2026 Agent: Production-Ready AI

Looking ahead, Parra predicts that "2026 is the year agents go to production." He elaborated on the vision for a 2026 agent, which will be capable of applying a wide range of skills, composing complex calls using MCP and CLI, connecting to various services, and running long-running tasks asynchronously. These agents will also feature rich interaction capabilities through MCP apps and discover tools progressively.

The key to achieving this is enabling agents to orchestrate their own actions. Instead of simply making one tool call per turn, models will write code that orchestrates multiple MCP tools, including loops, branches, and error handling. This programmatic approach allows for more sophisticated and adaptable agent behavior.

Key Challenges and Future Development

Parra also touched upon the challenges that still need to be addressed. He noted that current agents still "suck" in some areas, particularly in their ability to seamlessly integrate and orchestrate various tools and services. The success of future agents relies on:

• Progressive Discovery: Models need to efficiently discover and utilize tools. The comparison between "without tool search" (56,000+ tokens) and "with tool search" (~9,000 tokens) highlights the significant efficiency gains from effective tool discovery.
• Programmatic Tool Calling: Agents should be able to write code that orchestrates multiple tools, rather than just making sequential calls. This includes features like loops, branches, and error handling.
• Improved Core: Developing a more robust and scalable MCP, including stateless transport, improved tasks, and updated SDKs for various languages.
• Integrate Everywhere: Enabling cross-app access and server discovery via server cards to create a more interconnected agent ecosystem.
• Pushing the Boundary: Exploring experimental extensions, such as skills over MCP, to further enhance agent capabilities.

Parra concluded with a powerful statement: "2026 is all about connectivity. The best agents use every available method." This vision underscores the need for a flexible and interconnected protocol that allows agents to seamlessly integrate with a wide array of tools and services, ultimately leading to more powerful and versatile AI applications.