How a Multi‑Agent Team Built an HTML Page in One Take (No More “Continue” Prompts)

The author tried MiniMax’s newly released Agent product, Mavis, by giving it a simple prompt to create an HTML special‑topic page; after a 28‑minute nap the result was a fully rendered page with interactive elements, sidebars showing multiple agents, and a polished design.

What is an Agent Team?

The Agent Team consists of three role‑based agents:

Leader – coordinates the overall goal and decomposes the task into subtasks.

Worker – executes a specific subtask (e.g., content creation, design, HTML generation).

Verifier – inspects the output for factual accuracy, readability, and code correctness, then produces a verification report.

In the author’s experiment, Mavis acted as the Leader, while three Workers handled content, design, and programming respectively. The Verifier evaluated the final page, confirming that it contained the expected “star‑dust” background and particle effects.

Why single agents struggle

The article outlines three common failure modes of single‑agent workflows:

Agents stop midway and repeatedly ask “continue?” – the author spent half an evening repeatedly typing “continue”.

Long tasks cause the agent to become “dumber”, losing focus and mixing contexts, leading to repeated clarification requests.

In instant‑messaging interfaces the agent either replies with a shallow answer after 30 seconds or remains silent for minutes, creating a “cold‑war” experience.

“I already completed 1/2/3, need to continue?” – the agent’s habit of asking for confirmation.

How the Agent Team resolves these issues

The Leader first receives the high‑level request, breaks it into independent subtasks, and assigns each to a Worker. Workers run in parallel, each with its own isolated context, so research and coding steps do not interfere. The Verifier runs a separate check, ensuring quality without self‑bias. All coordination is governed by a state‑machine called the Team Engine , which defines explicit start, stop, retry, and verification conditions, removing the need for the model to guess when a task is finished.

This architecture eliminates the three failure modes: no more “continue” prompts, each Worker’s context stays focused (preventing the “getting dumber” effect), and the system proactively reports progress, so instant‑messaging windows never go silent.

The article also discusses token economics. While multi‑agent setups consume more tokens, the “Cost of Consensus” paper shows that in homogeneous debate settings token usage can be 2.1–3.4× higher without improving accuracy, highlighting the importance of a disciplined Team Engine to avoid waste.

Practical outcomes and future plans

Using the Agent Team, the author produced the HTML page in a single take, saving hours of iterative prompting and debugging. The system also supports mid‑task additions – the user can ask the Leader to spawn a new sub‑team, and the Leader replies with an immediate status update.

MiniMax plans to open‑source the Agent Team alongside the upcoming MiniMax M3 release. The desktop download is available at agent.minimaxi.com/download.

Overall, the multi‑agent approach demonstrates that architectural design, rather than simply scaling model size, is key to handling long‑running, complex tasks efficiently.