Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
Anthropic engineers argue single-agent conceptual integrity doesn't scale, requiring multi-agent orchestration
CJ Avilla and Harrison Stall from Anthropic open their AI Engineer talk by drawing parallels between AI agent scaling challenges and classic software engineering history — Ken Thompson's Unix (one person, conceptual integrity) versus IBM's OS/360 (1,000 engineers, coordination chaos) and Fred Brooks' Mythical Man-Month lessons. The framing sets up an argument that single-agent 'conceptual integrity' breaks down at scale, necessitating multi-agent architectures. The provided captions cut off before the technical content on running 1,000 agents, limiting the extractable signal.