Why Agents Need Connectivity, Not Better Models - David Soria Parra, Co-creator of MCP, Anthropic

Join us at AGNTCon + MCPCon NA 2026 https://hubs.la/Q04gbXLb0 David Soria Parra, co-creator of MCP at Anthropic, opens the conference with a case for where agentic AI is heading in 2026. This talk is for people building agents for knowledge workers rather than developers, and his argument is that the constraint has moved: model capability is compounding on its own, and what is left is connectivity. He traces why coding agents reached production first, why general knowledge work is next, and what the MCP project is building over the next twelve months to support it. - The long-horizon capability curve: models went from a few hours of unsupervised work to approaching multi-week tasks, on a straight line that shows no sign of bending. - Why 2024 was demos: tool calling was two years old, capabilities were flimsy, and nothing was production ready. - Why coding agents shipped first: compilers and unit tests could automatically steer and correct a model that could only work for a few hours at a stretch. - Why 2026 is different: as models get it right more often without steering, real production use cases open up in knowledge work, scientific discovery, and mathematics. - The knowledge worker problem: an accountant or finance advisor needs the model inside Microsoft 365 and Google Workspace, not a better model. - The adoption numbers: roughly one billion accumulated tool calls over MCP on Claude, over 500 million MCP SDK downloads month over month, and MCP servers as the fastest growing product line at companies like Vercel and Linear. - How the protocol evolved: remote connectivity, authorization, and applications added since the 2024 launch. - The stateless core: the biggest change to MCP in 18 months, built by Microsoft, Google, Anthropic, and OpenAI in a room together, including multi-round trip requests and cacheable results. - Agentic messaging and MCP Tasks: long-running operations that let you call an MCP server as an agent and get a result back minutes, hours, weeks, or months later. - Skills over MCP: serving agent skills through an MCP server, which makes a central company skill registry possible without building a marketplace. - Authorization and identity: the largest area of work in the MCP community, covering who an agent acted on behalf of, plus observability and policy. - Three ways to get involved in the MCP community, and why regional perspectives matter to the spec. If you are deciding whether to invest in MCP infrastructure for non-developer use cases, this is the framing to start from. LINKS AND RESOURCES - Model Context Protocol: https://modelcontextprotocol.io/ - The 2026-07-28 specification: https://blog.modelcontextprotocol.io/posts/2026-07-28/ - MCP on GitHub: https://github.com/modelcontextprotocol - Skills over MCP interest group notes: https://github.com/modelcontextprotocol/modelcontextprotocol/discussions/2248 - Agentic AI Foundation: https://aaif.io/ TIMESTAMPS 00:00 Setting the stage for 2026 and beyond 00:33 Why connectivity and open standards matter more than they already do 00:52 The long-horizon capability curve explained 01:45 Capabilities compound, and what that means for products 02:20 2024: tool calling was new and everything was a demo 02:55 2025: why coding agents became the first real production agents 03:35 Compilers and unit tests as automatic steering 04:05 2026: production agents beyond coding 04:35 What general knowledge workers need that developers do not 05:20 Connectivity as the real constraint 05:50 Adoption: a billion tool calls and 500 million SDK downloads 06:40 MCP across the industry, including competitors 07:10 How the protocol evolved since launch 07:40 Making MCP truly stateless, and who built it 08:20 Multi-round trip requests and cacheable results 09:00 What MCP needs for the next three years 09:25 Agentic messaging and MCP Tasks 10:10 Adding semantics: skills over MCP 10:55 A central skill registry without a marketplace 11:20 Authorization and identity 12:00 Three ways to get involved in the MCP community #ModelContextProtocol #AgenticAI #MCPTasks