The Protocol Stack for Production Agents: MCP, A2A, and What Comes Next
About this session
Every AI agent demo looks the same. A clean prompt, a happy path, and a round of applause. Production is nowhere near that tidy.
In this talk, I'll walk through the real architecture decisions behind building multi-agent systems with Google's Agent Development Kit, the Model Context Protocol (MCP), and the Agent-to-Agent (A2A) protocol, drawing on hands-on build work and enterprise MLOps experience. I'll cover where MCP servers earn their keep versus where they just add overhead, how to structure tool calls so agents behave reliably under real traffic, and the production issues that never make it into the tutorial: session handling, cost monitoring, context window management, and the temptation to let one agent do too much.
I'll also touch on agent skills governance and how teams should think about versioning, scoping, and validating agent capabilities (SKILL.md-style patterns) as agent systems scale beyond a single team's ownership.
You'll leave with a clearer mental model of the emerging agent protocol stack, a framework for scoping agent skills versus core application logic, and honest lessons on what broke first when moving from prototype to something closer to production.
Speaker
Key takeaways
- A working mental model for the MCP + A2A protocol stack and where each fits
- Practical patterns for structuring reliable agent tool calls, session handling, cost control, context management
- A framework for scoping agent skills versus core application logic and an introduction to agent skills governance