Your Agents Are in Solitary Confinement: Why MCP & A2A Aren't Enough — Vlad Luzin, Band

This talk argues that current methods for enabling AI agents to work together, such as running models side-by-side and manually routing messages, or using basic scripting, are insufficient. It identifies the 'single-agent bottleneck' and the limitations of human-centric messaging platforms (Slack, Discord) and low-level protocols (MCP, A2A) for bot-to-bot traffic. The core technical contribution is framing agent collaboration as a distributed systems problem, requiring specific architectural components: real-time ordered transport, persistence and rehydration, runtime binding across frameworks, agent-first abstractions (participants, rooms, routing), and enterprise governance. Two demos illustrate agents (Codex, LangGraph, Claude Code, personal assistant) discovering and collaborating in real-time.