Intel / video

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

Source: youtube.com
Author: AI Engineer
Date: Published Sep 30, 2026

Context

Getting several agents to collaborate requires a place for their work to meet. Band, which builds communication infrastructure, proposes shared rooms where people and agents work side by side while each agent keeps its own tools, models and memory. This addresses the manual relay involved when someone repeatedly carries one agent's plan to another agent for review.

In a conference talk, Band's co-founder and CTO Vlad Luzin, according to the talk's description, frames the remaining engineering work as real-time ordered messaging, persistence so agents can resume, and connecting different runtimes. He argues that protocols like and A2A are too low-level for this because they leave queues and timeouts to the builder. That is a claim by a company selling the alternative. The proposal supplies a coordination architecture, rather than evidence that adding more agents necessarily improves an answer.

Key takeaways · AI-distilled

Luzin's framing: running Claude and Codex side by side, one planning and one reviewing, makes you the network router passing messages by hand. Loop-engineering scripts automate that relay but still do not let the agents actually work together.
He argues bigger windows do not fix the single-agent bottleneck, and that Slack, Discord and WhatsApp are built for humans and block bot-to-bot traffic, so they cannot serve as the agent collaboration layer.
His specific complaint about MCP and A2A: stateless calls, one-way client/server roles and no discovery, which leaves developers to build their own queues and timeouts on top.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source www.youtube.com