Giving AI Agents a Voice in Live Team Conversations
Summary
High Fidelity describes an experiment putting agentic teammates into Quad, its browser-based spatial videoconferencing product, during team rituals such as user acceptance testing. One agent was manageable, but sessions with several agents became chaotic: they responded late, spoke after the discussion had moved on, volunteered unrequested information, and produced overly long answers. In one 65-minute session, Quad's clients requested that the server silence agents 61 times; the five-second cooldown plus 1.5 seconds of detected human speech meant that unwelcome audio could continue for roughly two seconds, compared with the few hundred milliseconds in which humans usually resolve overlapping speech. The company argues that ordinary turn detection does not solve the central problem. Turn detection asks whether a person has finished speaking, while a shared table needs a coordinator that decides which of several agents may speak and when the floor should be reclaimed. Quad therefore gave its table server floor control. Agents join through MCP and can be built on Claude or Codex, but all use the same table rules. When humans talk over an agent, the server purges queued speech and stops the agent; a 90-second quiet period then prevents unsolicited agent speech while allowing direct replies. The server records the reason for each speaking decision, with 23 reasons currently logged, including human control of the floor, another agent already answering, and a limit on agents responding to one another. The current design silences all agents at a table, does not enforce brevity, and allows an agent to hold the channel indefinitely. High Fidelity is considering separate controls, voice headlines with textual detail, social signals such as attention and hand-raising, quieter side-channel speech, and faster interruption handling. The article presents multi-agent voice coordination as an unresolved production problem for teams building shared live audio experiences.