The Future of MCP — David Soria Parra, Anthropic

Key Concepts

• MCP (Model Context Protocol): An open standard that allows AI agents to interface with external data, tools, and applications consistently across different platforms.
• Progressive Discovery: A pattern where an agent loads tools on-demand rather than dumping all available tools into the context window, reducing token usage and latency.
• Programmatic Tool Calling (Code Mode): Allowing the model to write and execute code (e.g., via a REPL or interpreter) to compose multiple tool calls into a single execution, rather than performing sequential, high-latency round trips.
• Stateless Transport Protocol: A proposed evolution of MCP to make servers easier to deploy on cloud infrastructure (e.g., Kubernetes, Cloud Run) by moving away from complex streamable HTTP.
• Cross-App Access: An upcoming feature for enterprise identity management, allowing users to authenticate once via providers like Okta or Google and access multiple MCP servers without re-logging.

---

1. The Evolution and State of MCP

The speaker highlights that MCP has grown from a local-only, tool-focused specification to a robust ecosystem with 110 million monthly downloads. This adoption rate significantly outpaces the growth of major open-source projects like React.

• Current Status: The ecosystem has moved beyond simple "toy" projects to complex SaaS integrations (Linear, Slack, Notion).
• The 2026 Vision: While 2024 was for demos and 2025 for coding agents, 2026 is defined as the year of production-grade general agents. These agents require deep connectivity to multiple enterprise systems rather than just local code compilation.

2. Connectivity Stack: Choosing the Right Tool

The speaker argues against a "one-size-fits-all" approach to agent connectivity. Instead, developers should utilize a stack based on the specific use case:

• Skills: Best for domain-specific knowledge and reusable capabilities.
• CLIs: Ideal for local coding agents, sandboxed environments, and tools already present in pre-training (e.g., Git, GitHub).
• MCP: Necessary when the agent requires rich semantics, UI rendering for long-running tasks, enterprise-grade authorization, governance, or platform independence.

3. Methodologies for Better Agent Harnesses

To improve agent performance, the speaker outlines three critical shifts in development:

A. Progressive Discovery

Instead of loading all tools into the context window, developers should implement a "tool-loading tool." The model should only fetch the specific tool definitions it needs at the moment of execution. This drastically reduces context window bloat.

B. Programmatic Tool Calling

Rather than having the model perform sequential tool calls (which is latency-sensitive and inefficient), developers should provide an execution environment (e.g., V8 isolate, Lua interpreter). The model writes a script to perform the task, which is then executed in one go. Using Structured Output allows the model to understand return types, enabling better composition of these tasks.

C. Designing for Agents, Not Just REST

The speaker criticizes the "cringe" practice of mapping REST APIs one-to-one to MCP servers. Instead, developers should:

• Design interfaces based on how a human would interact with the system.
• Use server-side execution environments (like the Cloudflare MCP server) to allow the model to orchestrate logic internally.
• Leverage MCP-specific features like Elicitation (asking the user for missing info) and Tasks.

4. Future Roadmap and Protocol Improvements

The speaker outlined several upcoming technical updates:

• Stateless Transport: A Google-led proposal to simplify deployment on hyperscalers.
• SDK Upgrades: Version 2 of the TypeScript and Python SDKs, incorporating lessons from the community-driven "Fast MCP" project.
• Server Discovery: A mechanism for agents/crawlers to automatically detect MCP servers via well-known URLs on websites.
• Skills over MCP: A new extension to ship domain knowledge alongside tools, allowing for continuous updates without needing complex plugin registries.

5. Notable Quotes

• "If someone tells you there's one solution to all your connectivity problems... they are probably pretty wrong."
• "2025 was all about coding agents... 2026 is all about putting these agents into production."
• "Every time I see someone building another REST-to-MCP server conversion tool, it's a bit cringe because it just results in horrible things."

Synthesis

The core takeaway is that the future of AI agents lies in seamless, multi-modal connectivity. Developers must move away from simple, inefficient tool-calling patterns toward a more sophisticated architecture that prioritizes progressive discovery, programmatic composition, and enterprise-ready protocols. By treating MCP as a "connective tissue" rather than just a plugin system, developers can build agents that are not only capable of coding but are also effective knowledge workers capable of navigating complex enterprise environments.