According to Dynamic Beating monitoring, MiniMax has overall upgraded its desktop Agent product and rebranded it as Mavis (MiniMax as a Jarvis). The core new capability is Agent Teams: users can create multiple Agents with different roles to form teams that collaborate to complete complex, long tasks that a single Agent may struggle with. At the same time, the previously separate API and Agent subscriptions have been merged into one, with CLI, API, and Agent all interconnected, sharing quotas.
MiniMax has also released a technical long read to explain the design concept of Agent Teams. This long read itself was generated by an Agent team—an Agent simulating user inquiries, and another one answering based on internal technical information. The article points out four core issues with single-Agent handling of long tasks: unexpected halts during task execution, deteriorating output quality due to extended context, blocking user real-time interaction during long tasks, and the inability to achieve true role separation at the prompt level.
To address these problems, Mavis employs a code state machine instead of prompt orchestration to drive collaboration. The team consists of three roles: the Owner is responsible for task breakdown and scheduling, the Worker focuses on execution, and the Verifier independently verifies the work. The Verifier and Worker form an adversarial mechanism, with strict isolation of contexts for each role, only communicating through structured summaries. MiniMax openly admits in the article that multi-Agent collaboration will introduce additional handover and aggregation costs. However, for tasks with long chains and high risks, this structured overhead results in the certainty of delivery outcomes.
