We started with Werewolf and Zoom — and ended up building a multi-agent workspace

When we first imagined Nexus, we weren’t trying to build another chatbot or a system that simply sends the same task to several AI agents. Our earliest inspiration came from two very different places: the social deduction game Werewolf and the idea of a shared Zoom meeting room. In Werewolf, every player has a distinct identity, while a moderator — sometimes called the “God” of the game — understands the overall structure, guides the players, and keeps the game moving. The players listen to one another, make decisions based on incomplete information, and act according to their own roles. That structure felt surprisingly suitable for AI collaboration: a human or host agent can coordinate the process, while several specialized agents participate with different identities, skills, responsibilities, and perspectives. However, we felt that most existing multi-agent systems were still missing something fundamental. They could divide a task among several agents, but those agents often behaved like isolated workers connected by a predefined workflow. They could produce separate results, yet they rarely communicated as freely as people do in a real team — listening to the same conversation, asking each other questions, handing work back and forth, or changing direction after hearing a better idea from another participant. This brought us back to the day-and-night structure of Werewolf. During the day, everyone participates in a shared discussion. The human can speak with the host, the host can coordinate individual agents, and agents can hear and respond to one another in the same public context. At night, selected players can communicate privately. In Nexus, this became directed communication between agents: one agent can privately ask another to verify a result, review some work, or prepare an answer without automatically exposing every intermediate message to the entire group. Once that private work is complete, the conclusion can be brought back into the public conversation. The Zoom metaphor gave this communication structure a natural home: the Room. A Nexus Room is a shared space where humans and multiple agents can discuss, delegate, and work together. Each agent also has its own Workspace, where its files, skills, memory, and long-term working context can persist. Public Rooms, private communication, and independent Workspaces together allow agents to collaborate more like members of a real team instead of temporary instances responding to isolated prompts. Our goal is not to create one all-knowing AI. We want to create an environment where humans and specialized agents can work together through public discussion, private coordination, clear roles, persistent context, and meaningful human control. What began as a combination of Werewolf and Zoom eventually became a larger question for us: what should communication look like when AI agents stop being tools we talk to one at a time and start becoming participants in a shared workspace? We’d love to hear what others working with multi-agent systems think: should agents be able to “whisper” to one another, and how much visibility or control should humans have over those conversations?