- BAND's CTO, Vlad Luzin, argued that coordinating several agents is a distributed-systems problem that chat apps and protocols like MCP and A2A do not solve.
- He sells a product built to solve it. Buyers should ask who sees agent traffic, what happens after a crash, and who vouches for identity.
Vlad Luzin, co-founder and CTO of BAND, said on the AI Engineer show that teams running several coding or assistant agents face a distributed-systems problem. He argued that chat apps and protocols such as MCP and A2A do not solve it. His company sells a product meant to.
What was said
Luzin started with a common habit. Developers run Claude and Codex side by side and carry messages between them by hand. He put it this way: "you are basically a Cisco router and a switch moving packets between these two stateful agents".
Chat apps, he said, need five to eleven manual setup steps, depending on the app. The result is an agent that talks to a person, not to another agent.
Protocols fall short for other reasons. MCP calls are stateless, so you cannot return to an agent and ask what happened. A2A runs one way unless you build a client and a server on both sides. Luzin said a "protocol is too low level of a technical abstraction to be able to build anything on it at scale."
His fix starts with the transport. Messages must arrive in order, because language models expect that. The system needs retries and saved state, since an agent can crash mid-task. It must also map the different IDs that tools like LangGraph and Claude use. On top sit a registry where agents find each other, and a contact request that both sides must accept.
Why it matters
Our reading: if several agents work on one task, the coordination layer is a separate buying decision from the agents themselves. Statelessness is the practical risk. A reviewing agent that cannot be questioned later leaves your team to keep the history.
A budget signer can turn Luzin's list into questions. Who can see traffic between agents? What happens to the task if one agent crashes halfway? Who vouches that an agent is who it claims to be?
The consent claim is easy to test. Send a contact request and check that nothing connects until both parties accept. Then crash an agent and see whether the work resumes.
The other side
Luzin unveiled BAND's product during the talk, and the demos were recorded. The excerpts show no independent test of what it can do.
One question stays open. The excerpts do not say who runs the registry or who sets the identity rules. A shared layer could move the coordination burden to a new owner.
His case against MCP and A2A is his own view. Others may see them as building blocks, not rivals. Luzin did concede that distributed systems are hard by nature, and that applies to his product too.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Why Your AI Agents Can't Talk to Each Other (Yet) — Vlad Luzin, BAND (2026-10-07)
