Reborn as a Boss in the AI Era: Let a Team of Agents PUA Each Other

MiniMax has unveiled a new agent product called Mavis, which stands for MiniMax as a Jarvis. The name hints at its ambition: to act as a personal assistant that can handle complex tasks autonomously. However, what sets Mavis apart is not just its individual capabilities, but its ability to form a team of agents that work together.

In a recent demonstration, a user gave Mavis a single, vague instruction: "Create an HTML feature page based on Mavis's blog." Without any further guidance, Mavis spent 28 minutes planning, executing, and delivering a fully interactive HTML page with animations and a starfield background. The user, who had taken a nap during the process, was surprised to find the task completed without any prompts for continuation or clarification.

The secret lies in Mavis's agent team structure. When given a task, Mavis acts as a Leader, decomposing the overarching goal into sub-tasks. It then assigns these sub-tasks to specialized Worker agents—in this case, a content creator, a designer, and a programmer. A Verifier agent later checks the work for factual accuracy, readability, and code functionality, generating a report. This Leader-Worker-Verifier setup resembles a corporate team, with the user acting as the CEO who only needs to communicate with the leader.

This approach addresses three common frustrations with single-agent systems:

1. Constant need for confirmation: Single agents often stop mid-task and ask "Should I continue?" Mavis's Team Engine, a state machine, controls when to stop, retry, or move on, eliminating the need for constant user input.
2. Performance degradation on long tasks: As context grows, single agents tend to lose focus. In Mavis, each Worker operates in an isolated context, preventing interference between tasks like research and coding.
3. Silent IM responses: When a single agent is busy, it may not respond to messages. Mavis's main agent immediately acknowledges the task, while sub-tasks run in the background. Users can even add new requirements mid-execution, and the leader will adjust accordingly.

The architecture also introduces an adversarial relationship between Worker and Verifier. The Worker stops when the Verifier finds issues, and the Verifier stops when it can no longer find problems, driving iterative improvement. This mirrors the dynamic between R&D and quality assurance in a company, reducing the need for human oversight.

Cost is a concern with multi-agent systems, as token consumption is higher. However, MiniMax has addressed this by merging TokenPlan and Agent Plan subscriptions, allowing shared credits between API and agent use. The company also emphasizes that the Team Engine can decide when a single agent suffices versus when a full team is needed, preventing waste. A research paper cited in the article, "Cost of Consensus," notes that unstructured multi-agent systems can be 2.1 to 3.4 times more expensive than single-agent self-correction without accuracy gains. MiniMax's structured approach aims to avoid this pitfall.

Looking ahead, MiniMax plans to open-source the Agent Team framework, likely alongside the release of MiniMax M3. The company envisions users managing agent roles, capabilities, and boundaries through a control panel, shifting the focus from prompt engineering to team management.

As the article concludes, "Team was never the default option." For simple tasks, a single agent or even a script may suffice. But when complexity demands it, a well-organized team of agents can outperform a solo agent, handling long-term tasks with minimal intervention. The era of multi-agent collaboration has arrived, and with it, a new role for humans: not just users, but managers of AI teams.