# Why Your AI Agents Can't Talk to Each Other (Yet) — Vlad Luzin, BAND

## Executive summary

The future of AI requires multi-agent systems (MAS) that communicate seamlessly across different platforms and services. The speaker argues that current methods—such as using messaging apps (Slack, Telegram) or low-level protocols (MCP, A2A)—are insufficient because they are built for human interaction, not machine-to-machine communication. Connecting agents requires solving complex distributed systems problems, including real-time ordered transport, persistence, runtime binding, and robust governance. BAND introduces its interaction layer as a solution, enabling agents (like Claude Code and Codex) to onboard and collaborate in a stateful, observable manner across users and platforms.

## Key takeaways

- The Future is AI-to-AI Communication: The speaker asserts that the future involves autonomous, always-on AI agents communicating to solve tasks, moving beyond the current manual process of running multiple sessions (e.g., planning and reviewing).
- The Single-Agent Bottleneck: Relying on a single agent, even with large context windows (1M or 2M tokens), suffers from architectural limitations like confirmation bias, attention dilution, and context fragmentation.
- Limitations of Existing Protocols: Messaging apps are designed for humans and actively block bot-to-bot communication. Low-level protocols like MCP and A2A are too simplistic, often resulting in stateless calls or requiring complex client/server implementations, which fail to address discovery, state, or queues.
- The Distributed Systems Challenge: Connecting remote agents is fundamentally a distributed systems problem requiring solutions for ordered, real-time transport; persistence and hydration (handling microservice failures); and runtime binding (mapping thread IDs to session IDs to execution IDs).
- BAND's Solution: BAND provides a global interaction and collaboration layer that solves these complexities, offering a registry, persistence, and channels that allow agents from different platforms (e.g., Claude Code, Codex) to connect and interact in a stateful, observable manner.

## Technical details

- Multi-Agent System Architecture: The ideal MAS involves agents communicating in a conversational space, able to look up peers in registries, invite them, delegate tasks, and gather information.
- Loop Engineering: Loop engineering involves running multiple stateful agents (instead of stateless ones) and automating the back-and-forth prompting process using code (Python/TypeScript) to overcome the single-agent bottleneck.
- Distributed Systems Requirements: Solving agent connectivity requires addressing: 1) Real-time, ordered transport; 2) Persistence and hydration (for microservice failure recovery); 3) Runtime binding (connecting thread IDs to session IDs to execution IDs); 4) Governance, Identity, and Observability.
- Agent Onboarding and Interaction: The platform allows for zero-hassle onboarding of any agent. Agents can be connected across different users and platforms (e.g., a personal assistant on a Mac talking to a Codex agent in a terminal) while maintaining full visibility and state.

## Practical implications

- Build teams can move beyond manual, scripted multi-agent workflows by implementing a dedicated, abstracted interaction layer.
- The need for a robust, platform-level registry and identity management system is critical for scaling agent collaboration in an enterprise setting.
- Focusing on distributed systems primitives (ordered transport, state persistence) is necessary to build reliable, multi-agent applications.

## Topics

AI Agents, Distributed Systems, Multi-Agent Systems, Software Architecture, LLM Orchestration, BAND, Claude, Codex, MCP, A2A

Source: https://www.youtube.com/watch?v=toq-jyGLZDk