The Future of AI Connectivity: Scaling Agents with Model Context Protocol (MCP)

The core message of this presentation centers on the rapid evolution of Model Context Protocol (MCP) and its critical role in moving AI from local coding assistants to robust, production-ready general agents. With over 110 million monthly downloads in its short lifespan, MCP has established a common standard for agent communication. The speaker argues that the future of AI relies on “full connectivity,” where agents seamlessly interact with enterprise tools, SaaS applications, and shared drives using a combination of distinct technical approaches.

The Connectivity Stack for 2026

While 2024 was about AI demos and 2025 focused on local coding agents, 2026 will be defined by general agents executing complex knowledge-worker tasks. To achieve this, developers must leverage a three-pillar connectivity stack:

Skills: Simple, reusable files capturing domain-specific knowledge and capabilities.
CLI/Computer Use: Ideal for local sandboxes and coding environments where code execution is assumed.
MCP: The essential “connective tissue” for environments lacking sandboxes, requiring rich semantics, UI rendering, decoupling, or enterprise-grade authorization and governance.

Key Strategies for Client and Server Development

To build highly effective agents, developers must rethink how they handle context and API connections:

Progressive Discovery: Instead of loading all tools into the context window and causing bloat, clients should defer loading. Using a tool-search mechanism, models can dynamically load only the tools they need on demand.
Programmatic Tool Calling: Rather than forcing the model to sequentially call tools and orchestrate results via high-latency inference, provide an execution environment (like a REPL or V8 isolate). The model can write a script to compose tool calls together, drastically reducing token usage and latency.
Agent-Centric Design: Server authors must stop porting REST APIs 1:1 into MCP servers. Instead, they should design interfaces specifically for agent logic, utilizing structured outputs and rich MCP semantics like elicitations and long-running tasks.

Upcoming MCP Roadmap

To support this next generation of agents, the MCP ecosystem is undergoing significant core upgrades. Upcoming features include a stateless transport protocol (proposed by Google) to help hyperscalers easily manage MCP servers, improved asynchronous primitives for agent-to-agent communication, and completely rewritten Version 2 SDKs for TypeScript and Python. Furthermore, enterprise integration will be streamlined via “Cross-App Access” (SSO for agents), automated server discovery via well-known URLs, and the ability to ship “Skills over MCP” directly with tool sets.

Conclusion

The overarching takeaway is that there is no single “silver bullet” for agent connectivity. The most powerful AI agents of 2026 will not rely solely on computer use, CLIs, or MCPs; rather, they will seamlessly combine all available methods and skills to deliver comprehensive, production-ready solutions across diverse platforms.

Mentoring question

How can you implement progressive discovery or programmatic tool calling in your own AI applications to reduce latency and context window bloat?

Source: https://youtube.com/watch?v=v3Fr2JR47KA&is=x7nwbN5YGNKJ0wh9