MiniMax has overhauled its desktop agent product, rebranding it as Mavis (short for MiniMax as a Jarvis) and introducing a new capability called Agent Teams. The feature lets users create multiple agents with distinct roles that work together to tackle complex, long-running tasks that a single agent couldn't handle. At the same time, the company merged its previously separate API and Agent subscriptions into a single plan, with shared usage quotas across CLI, API, and Agent.
To explain the thinking behind Agent Teams, MiniMax published a lengthy technical article — which was itself generated by an Agent Team. One agent simulated a user asking questions, while another answered based on internal technical documentation. The article identifies four core problems with single-agent long tasks: unexpected interruptions mid-execution, degraded output quality from overly long context windows, long tasks blocking real-time user interaction, and the inability of prompt-level roleplay to achieve genuine division of labor.
Mavis addresses these issues with a code-based state machine rather than prompt-based orchestration. The team consists of three roles: an Owner handles task decomposition and scheduling, a Worker focuses on execution, and a Verifier independently checks results. The Verifier and Worker operate in an adversarial dynamic, with each role's context strictly isolated and communication happening only through structured summaries. MiniMax acknowledges that multi-agent collaboration introduces additional overhead for handoffs and aggregation, but for long, high-stakes tasks, that structural cost buys reliability in the final output.