MCP in the Wild: Real-Time Agentic Systems with Live A/B Testing

Ever built a game where AI agents make decisions in real time—and talk to each other in a structured way? With the Multi-Context Protocol (MCP), you can do precisely that.
In this session, we’ll walk through how MCP enables you to split agent behavior into modular contexts, such as planning, reasoning, and execution, then wire them up into a real-time multiplayer game using Python, CrewAI, and FastAPI. Each agent runs its logic, and here’s the kicker: the underlying LLMs are hot-swappable. Swap in GPT-4, Claude, Mixtral—you’ll see the differences instantly, without touching your core code.
We’ll also demonstrate how to run A/B tests across agents by swapping models mid-game and capturing structured MCP metrics, including the time agents spend planning, the frequency of replanning, and the impact of context-switching on outcomes. These insights help you benchmark LLM performance beyond just token count or latency—and tune your architecture accordingly.

We’ll go hands-on with the open source repository, break down the architecture, and demonstrate how this protocol-based design can enhance not just games, but any multi-agent system. If you’re building with LLMs and tired of spaghetti orchestration, this talk is for you.