Google Cloud’s agent-teamwork experiment used a deliberately awkward domain: autonomous film crews, not just code. That makes the lessons more interesting. The post describes teams of agents with distinct roles, shared files, messages, verification gates, and a coordinator running multiple teams through waves.
The practical insight is not that agents can make media. It is that agent teamwork gets better when coordination is externalized. Teams that recorded decisions in shared files recovered from crashes and context loss. Teams that kept decisions inside message history lost direction when agents restarted.
The role design matters too. Idea, technical, editor, and coach roles forced different kinds of judgment. A coach could intervene only at step boundaries, which made verification a gate rather than micromanagement. That is closer to a production workflow than a swarm of unconstrained chat agents.
For New Runtime, the same pattern applies to editorial agents. Hermes, Codex, and OpenClaw should coordinate through files, manifests, receipts, and review gates. The chat is an interface; the durable workflow is the artifact graph.
New Runtime synthesis: multi-agent systems need shared state outside the conversation. Files, roles, and gates are what turn parallel model instances into a recoverable team.