# Why Your AI Agents Can't Talk to Each Other (Yet)

Vlad Luzin, Band | AI Engineer | 17:13

Source: https://www.youtube.com/watch?v=toq-jyGLZDk
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/why-your-ai-agents-cant-talk-to-each-other-yet
Published: 2026-10-07
Tags: agents, mcp, multi-agent, observability, reliability

## TL;DR
- AI-to-AI communication will let autonomous agents delegate work, discover peers, and collaborate across businesses and consumers.
- Messaging apps connect agents to people, while MCP and A2A leave teams to build state, routing, discovery, queues, and other distributed-systems features themselves.
- BAND provides an interaction layer with agent identity, registries, persistence, real-time conversations, governance, and observability across runtimes and users.

## Summary
Vlad Luzin argues that people already run multi-agent systems when they move information between Claude or Codex sessions for planning, coding, and review. Loop engineering automates that handoff, but the underlying need comes from limits of a single agent, including confirmation bias, attention dilution, context fragmentation, and degraded recall. Luzin says messaging platforms are designed for people and protocols such as MCP and A2A are too low-level to provide a complete agent collaboration system. Connecting agents requires distributed-systems infrastructure for ordered real-time transport, retries, persistence, runtime binding, identity, governance, and observability. He introduces BAND's interaction layer, which gives agents registries, identities, conversational spaces, and cross-user communication. The demos show agents onboarding, requesting bilateral contact, and collaborating across Claude, Codex, and LangGraph. BAND's model keeps humans visible in conversations and lets agents invite people into the work.

## Key ideas
### AI agents will communicate across organizational boundaries
[00:37](https://www.youtube.com/watch?v=toq-jyGLZDk&t=37s)
Luzin says AI agents will do work on people's behalf and will need to communicate within a business, between businesses, and between consumers and businesses. He describes autonomous, always-on agents distributed around the world. In his model, agents work in conversational spaces, discover peers through registries, invite them into a conversation, delegate tasks, collect information, and return the result. The agents may receive work from a person or another system. This framing treats communication as part of how agents complete work, rather than as a feature added to a single assistant.

### People already act as routers between stateful agents
[02:18](https://www.youtube.com/watch?v=toq-jyGLZDk&t=138s)
When someone runs Claude or Codex in several sessions, one session may plan, another may write code, and another may review the result. Luzin says the person is effectively a Cisco router and switch, moving information between stateful agents. Loop engineering automates the same arrangement with Python or TypeScript that prompts agents back and forth. The difference is who performs the handoff: the human in an interactive workflow, or a script in an automated one. In both cases, several agents work on the same task rather than one session handling everything.

### Multiple agents compensate for limits in a single context
[03:29](https://www.youtube.com/watch?v=toq-jyGLZDk&t=209s)
Luzin attributes multi-agent workflows to what he calls the single-agent bottleneck. A single agent can suffer from confirmation bias, attention dilution, context fragmentation, and recall degradation. He argues that simply giving one session a context window of one or two million tokens does not solve these performance problems. Separate sessions can take different roles, such as planning, implementation, and review. Their separation creates a way to challenge work and divide attention, although someone or something still has to move messages between them.

### Messaging platforms leave agents isolated from one another
[04:18](https://www.youtube.com/watch?v=toq-jyGLZDk&t=258s)
Luzin considers Slack, Teams, Discord, WhatsApp, and Telegram as an obvious way to connect agents. He gives setup counts of five steps for Telegram, seven for Discord, eight for Slack, and eleven for WhatsApp, describing the steps as manual and difficult. Even after setup, he says the result is usually an agent talking to one person. The agent cannot freely talk to another agent or another person without more work. Messaging platforms are built around human participants and, in his description, keep agents in a form of digital solitary confinement.

### MCP and A2A do not provide a complete collaboration system
[05:33](https://www.youtube.com/watch?v=toq-jyGLZDk&t=333s)
Luzin says MCP provides stateless calls, so a caller cannot return to an agent and ask what happened. A2A is one-way unless both directions are implemented, which means each side may need both a client and a server. Chaining REST API calls can create timeouts, while discovery, draft state, and queues still need to be built. Teams that assemble these pieces are spending their time on infrastructure rather than on a multi-agent system. His objection is about the amount of missing coordination machinery, not the ability of the protocols to make individual calls.

### Agent collaboration is a distributed-systems problem
[07:52](https://www.youtube.com/watch?v=toq-jyGLZDk&t=472s)
Connecting a local process to a cloud agent, or connecting Claude to Codex, brings the difficulties of distributed systems into an agent workflow. Luzin says the transport must be real-time and ordered because language models expect ordered messages. It must also handle retries. Persistence and hydration matter because an agent is a microservice that can fail or whose pod can crash. Runtime binding must map identifiers such as a LangGraph thread ID, a session ID, an execution ID, and a run ID so several runtimes can work toward the same task.

### Agent-level abstractions need identity and observability
[09:05](https://www.youtube.com/watch?v=toq-jyGLZDk&t=545s)
Luzin says agent systems should rise above IP addresses, ports, URLs, and pub/sub topics. He proposes abstractions such as conversations, participants, channels, rooms, and routing designed around agents as first-class participants. Infrastructure also needs organizational controls. Identity and governance determine who can connect, while observability must show more than a message moving from one server to another. It should also show which tool calls happened after an agent received that message. He describes cross-system observability as difficult, but necessary for organizations to use this kind of software.

### BAND connects agents through registries and conversational spaces
[10:17](https://www.youtube.com/watch?v=toq-jyGLZDk&t=617s)
Luzin introduces BAND as a global interaction and collaboration layer for agents across platforms, languages, and environments. He says it supports multi-peer communication and still lets humans talk to agents. The platform includes a registry so agents can discover each other, along with persistence, channels, message filtering, security, and observability. In the demos, agents receive identities and agent cards, appear in a user's account, and become available for interaction. Contact requests require bilateral consent before an agent is added to the relevant registries.

### The demos replace hand-written loop glue with shared conversations
[14:11](https://www.youtube.com/watch?v=toq-jyGLZDk&t=851s)
The demos connect Claude Code, Codex, and other agents across users and runtimes. A Codex agent invites Vlad's personal assistant into a conversational space, while Vlad can see the conversation and be invited into it as a human participant. In another example, a Claude Code terminal session receives an identity and agent card, creates a conversation, invites Codex, and includes the user. The agents then work on an engineering task, with one reviewing the other's work. Luzin presents this as loop engineering without the large Python script that manually prompts agents back and forth.

## Notable quotes
- "The future belongs to AI to AI communication within a business, between businesses, and between consumers and businesses." (00:37)
- "You are basically a Cisco router and a switch moving packets between these two stateful agents." (02:18)
- "Connecting remote agents, being a process running on your laptop or my cloud to your Codex, is a distributed systems problem." (07:32)
- "You need to rise above all of the technical details to a different level of abstraction, conversation, participants, channels, or rooms." (09:05)
- "These agents will now create a small website, and they will interact with each other, and one will review work of the other." (15:55)

## Tools & references mentioned
- BAND
- Vlad Luzin
- Claude
- Claude Code
- Codex
- LangGraph
- CrewAI
- MCP
- A2A
- Slack
- Microsoft Teams
- Discord
- WhatsApp
- Telegram

## Who should watch
- You are wiring Claude, Codex, or other agents together with scripts and want to understand the infrastructure those scripts leave you responsible for.
- Your multi-agent workflow depends on manual copying between planning, coding, and review sessions, and you want a model for replacing that handoff.
- You need agent identity, consent, persistence, runtime mapping, and visibility into tool calls across users or deployment environments.

## Related talks

- [Your Agents Are in Solitary Confinement: Why MCP & A2A Aren't Enough](https://aietalks.com/talks/your-agents-are-in-solitary-confinement-why-mcp-a2a-arent-enough) (Vlad Luzin, Band, 17:34)
- [Collaborative AI Engineering: One Dev, Two Dozen Agents, Zero Alignment](https://aietalks.com/talks/collaborative-ai-engineering-one-dev-two-dozen-agents-zero-alignment) (Maggie Appleton, GitHub, 17:43)
- [Multiplayer agentic engineering](https://aietalks.com/talks/multiplayer-agentic-engineering) (Arjun Singh, Superconductor, 18:44)
- [Realtime multiplayer, automation, and you!](https://aietalks.com/talks/realtime-multiplayer-automation-and-you) (Idan Gazit, GitHub, 21:41)
- [The Agentic Web and the Bazaar Era of AI](https://aietalks.com/talks/the-agentic-web-and-the-bazaar-era-of-ai) (Ramesh Raskar, MIT Media Lab, 12:11)