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

## Executive summary

Multi-agent coordination is fundamentally a distributed systems problem, not a simple scripting task. Current methods—including low-level protocols like MCP and A2A, or messaging platforms like Slack—are insufficient because they lack necessary features such as real-time ordered transport, persistence, discovery, and runtime binding across diverse frameworks. The speaker introduces Band, a global interaction layer designed to abstract these complexities, enabling autonomous agents (like Codex and LangGraph) to communicate and collaborate seamlessly regardless of their underlying framework or deployment environment.

## Key takeaways

- The Multi-Agent Problem: Attempting to coordinate agents by hand (e.g., copy-pasting between two sessions) or using basic protocols (MCP/A2A) requires the developer to act as a 'router,' which is inefficient and fails to manage state, persistence, or discovery. (2:04, 5:33)
- Limitations of Existing Protocols: Protocols like MCP are completely stateless, and A2A is client-server. Chaining multiple agents involves managing REST API timeouts, and developers must manually implement queues, persistence, and discovery mechanisms. (5:33)
- The Distributed Systems Requirement: For agents to work together reliably, the system must solve the transport layer (ordered, real-time delivery, retries), continuity (handling crashes/rehydration), and runtime binding across different agentic frameworks. (8:17)
- The Solution: Band: Band provides a global collaboration layer that abstracts the technical stack to concepts like 'rooms,' 'channels,' and 'participants.' It handles the necessary primitives for agent-to-agent communication, allowing agents from different frameworks (e.g., Codex, LangGraph) to connect and collaborate autonomously. (10:17)

## Technical details

- Agent Communication Architecture: The future requires agents to communicate autonomously, forming conversational spaces where they can discover each other and solve tasks without human intervention. (0:00)
- Protocol Deficiencies: Protocols like MCP and A2A are criticized for being too low-level, lacking built-in state management, discovery, and robust handling of message queues and timeouts. (5:33)
- Required Abstractions: A robust solution must provide ordered message delivery, consistency hydration, and runtime binding across diverse agentic frameworks. (8:17)
- Observability and Governance: The platform must provide full visibility into multi-agent work, including tracking token usage, cost attribution, and identifying which human or agent was involved in a specific task (user chain cost). (14:11)

## Practical implications

- Reduces the need for developers to manually code complex message routing and state management between agents.
- Enables seamless collaboration between agents deployed in different frameworks (e.g., connecting Salesforce, Slack, and Databricks agents).
- Provides crucial operational visibility for managers, tracking cost attribution and human involvement across autonomous agent workflows.

## Topics

Multi-Agent Systems (MAS), Distributed Systems, AI Orchestration, Agent Communication Protocols, Band, Codex, LangGraph, Claude Code

Source: https://www.youtube.com/watch?v=UOcHfR3_tys