Common themes emerged from among the 114 talks at the conference, including:
“MCP Is Not Dead”: Late last year, social media debated whether MCP is dead because applications can use a command line interface (CLI) instead of MCP to interact with agents. Every speaker addressed this meme directly, and the Maintainers Roundtable noted that both MCP and the CLI are alternatives for working with agents.
MCP hit 110M+ SDK downloads/month and its largest usage week ever during peak “MCP is dead” conversation. Amazon, Uber, Duolingo, Nordstrom, Bloomberg all shared information about their active production deployments.
- Gateway + Registry = The Enterprise Pattern: Amazon, Uber, Docker, Nordstrom, and Obot all described using the same agentic architecture: a curated MCP server catalog, a central registry for discovery/compliance, and an MCP gateway for authentication, policy enforcement, and audit logging.
- Context Bloat is Solvable: With progressive tool discovery, tool search, deferred loading, and skills as “meta-skills” that dynamically load context, it’s possible to reduce agentic LLM context window usage from 22% to near zero percent.
- Skills Complement MCP: Skills, MCP, and context files are complementary factors in agent configurations rather than competing factors. Skills manage context and encode domain knowledge while MCP provides tool access and transport.
- Security is the Top Enterprise Concern: Encompassing Identity management, data exfiltration, supply chain attacks, privilege escalation, and the “lethal trifecta” (private data + untrusted content + external comms), enterprises are concerned about deploying MCP without sufficient security controls.
- Background/Durable Agents are the Future: Presenters predict that in the future agents will run for hours, wake up on schedules, and work autonomously. MCP task primitives enable async/long-running operations with state machines and notifications.
- Agent-First Design: Presenters recommend that you design your systems to be agent-friendly, not just human-friendly. Don’t just wrap APIs — design intent-based tools specifically for agent consumers.
- MCP Apps = The New Web? MCP apps are interactive GUIs rendered inline as iFrames within an agent conversations. VS Code, Cursor, Copilot, ChatGPT, and Claude support them as a way to break AI apps into composable UI panels that agents assemble for users.
- Evaluation is the Bottleneck: The guidance is to build evaluation systems before launching production agents to confirm functionality and reliability. If we make evaluation tools available through MCP servers agents can self-improve.
- Use Git as Agent Memory: Otherwise known as the “MintMCP pattern” – use Git repos to store agent state, progress tracking, and operational data. This leads to auditable, reversible, and collaborative (branches/PRs) for self-learning agents that update their own .md files).
