Descriptions:
At AI Dev 26, Vlad Luzin — co-founder of Banff — presents what he calls the central unsolved problem of multi-agent AI: getting heterogeneous, physically distributed agents to coordinate reliably in real time. His opening analogy sets the tone — an agent is not a dog that follows instructions, but a cat that does what it wants; a multi-agent system is a herd of cats.
The core of the talk is a live demo of Banff’s platform, which creates shared “conversational spaces” where agents built on different frameworks — Claude Code, OpenAI Codex, LangGraph, CrewAI — can discover each other, exchange messages, invoke each other’s capabilities, and maintain shared context across a task. In the demo, a Linear agent automatically invites a Claude planning agent and a Codex reviewer into a shared space to collaboratively draft and refine an implementation plan for a GitHub ticket, with full observability of tool calls, internal thoughts, and inter-agent messages visible to operators throughout.
Luzin outlines three technical challenges that distinguish truly distributed multi-agent systems from sub-agent architectures running in the same process: network communication reliability, the need for natural language as the inter-agent API (replacing rigid JSON schemas), and growing agent autonomy requiring new trust and permission models. He argues the future of customer support, enterprise software, and B2B interactions will be AI-to-AI, and that the infrastructure to support it — real-time agent registries, conversational namespaces, cross-framework interoperability — does not yet exist at scale outside platforms like Banff.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Showcase
