A conference talk by an AI product manager at Koda tracing the evolution of their AI code review product from a single LLM-based agent to a multi-agent system. The journey covers early context window limitations requiring data pruning, adoption of MCP for dynamic context collection, a custom MCP transport layer built to avoid Node.js dependency issues in JetBrains plugins, and the progression from opinionated agents to ReAct agents. The talk culminates in a 'mixture of agents' architecture for code review: a diamond-shaped flow with a context collector feeding specialized parallel sub-agents (bug finding, compliance enforcement) whose outputs are filtered by an LLM-as-judge. LangGraph is used for agent orchestration with PostgreSQL for state management.