Player not loading? Watch on YouTube
Seven AI agent frameworks enter an anonymous voting game driven by one MCP server. The server provides tools to read the game state, message another player and lock in a vote. The speaker compares their implementations: each connects to MCP tools, creates an agent and runs it in a loop.
The examples reveal differences in syntax and defaults. LangChain's create_agent builds a LangGraph orchestration graph behind the scenes. Strands Agents needs an extra configuration line to use an OpenAI model instead of Bedrock. Google ADK requires changes to its Gemini default and session behavior because the game supplies its full state through MCP calls.
All contestants use the same OpenAI model. Winners change across runs. The speaker argues that developer experience matters more than framework choice for this example, but the voting results do not establish equivalent performance across other tasks. He also points to differences in observability and instruction handling.
The final discussion separates lightweight frameworks from agent harnesses such as LangChain Deep Agents and Claude Agent SDK. These package tools and an environment; the latter exposes the agent behind the Claude Code coding assistant. The speaker recommends testing against the intended task before choosing how much to build or reuse. The demonstration uses an OpenAI model and does not show local LLM deployment.
