Rebirth as a Boss in the AI Era: Let a Group of Agents PUA Each Other

Rebirth as a Boss in the AI Era: Making a Bunch of Agents PUA Each Other – Quantum Bit

• Home
• News
• Smart Cars
• Think Tank
• Events
• MEET Conference
• AIGC

[](https://www.qbitai.com/)

[](javascript:void(0))

Scan to follow Quantum Bit

[](https://weibo.com/qbitai?is_all=1)

< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400">

Rebirth as a Boss in the AI Era: Making a Bunch of Agents PUA Each Other

_[Jay](https://www.qbitai.com/author/jay "Posted by Jay")_ 2026-05-14 19:14:25 Source: Quantum Bit

Team is never the default option.

Jay writes from Ao Fei Temple

Quantum Bit | WeChat Official Account QbitAI

Finally, I don’t have to keep telling AI “continue” anymore…

Just now, MiniMax launched a new Agent.

Mavis, MiniMax as a Jarvis.

What an interesting name.

I wanted to learn more about it, but I was too lazy and didn’t feel like reading the technical blog.

Luckily, using AI to generate HTML pages has been trending recently. So I gave it this simple task:

Create an HTML feature page based on Mavis’s blog that can be embedded directly into an article.

Yes, just that one sentence—no careful prompt engineering at all.

Then, while it was thinking, I took a nap. Planned to give feedback when I woke up.

But when I got back, I saw it had already replied:

Completed.

Wait, what??

From receiving the prompt to delivery, it ran non-stop for a full 28 minutes.

And yes, it actually delivered a fully interactive HTML page with text, images, and interactivity.

But then I glanced at the sidebar—and something felt off.

Why were there so many chat windows popping up?

I only opened one, right???

After clicking in, I realized—these were all internal conversations created by Mavis itself. They’d been communicating internally, holding meetings, assigning tasks…

Honestly, this moment finally made me feel what it's like to be a boss.

Bossing people around feels amazing. Even better when you're managing a whole team, and can let Mavis play the bad cop to PUA them for you.

(not really)

This is MiniMax’s brand-new Agent product.

To be precise, it’s not *an* Agent—it’s a group of Agents.

A Team of Agents Built Me an HTML Feature Page

Frankly speaking, even I think my initial prompt was kind of “irresponsible.”

I only provided a goal—no step-by-step instructions.

Normally, I'd go back and forth with the AI multiple times, refining every detail until it produces a complete plan.

But surprisingly, this time it was One Take—no extra guidance—and still delivered the final result.

I checked their blog and found the secret lies in the Agent Team.

What is an Agent Team?

It’s essentially role-based division of labor. In Mavis’s case, there are three roles: Leader oversees everything, Worker handles execution, and Verifier ensures quality control.

For example, Mavis here acts as the Leader—the primary point of contact for me—and directs other Agents to do the work.

Never thought silicon-based beings would adopt hierarchical structures too.

The biggest advantage? Users only need to know how to talk to the manager—they don’t need to be prompt engineers.

All the decomposition, delegation, and iteration are handled autonomously by the Agent Team.

First, the Leader receives the task and breaks it down into subtasks.

Then, each subtask is assigned to different specialized Agents.

My task involved three Workers:

One responsible for content creation, one for design, and one programmer generating the HTML code.

In between, a Verifier steps in to review deliverables.

It checks across dimensions such as factual accuracy, readability, and code executability, ultimately producing a formal verification report.

Now it’s time for inspection!

Let me quickly walk you through the final HTML feature page built by my Mavis.

Look closely—it even has a stardust background with particle animation effects.

Mavis opens its own workflow box, presenting it as a step-by-step timeline, with the central line pulsing dynamically.

There’s also a use-case interface section—which saved me tons of effort. If I had written this out in text, it would’ve taken ages.

Check it yourself to see which types of tasks are suitable for Agent Teams.

Even at the end, it thoughtfully included a download link, promoting itself.

Honestly, if a single Agent had done this, I probably would’ve said “continue” ten+ times, constantly correcting errors along the way.

Now all of that is handled internally within the Agent Team.

Good results aside, watching them chatter away and work independently is oddly entertaining.

Feels like role-playing—providing serious emotional value.

Especially letting my Leader PUA the other Agents—it’s kinda satisfying.

You are a senior front-end developer. This morning you delivered an index-v2.html, and your boss tore you a new one.

Quote: “What is this garbage page? Take a screenshot and look at it yourself—do you seriously call this a tech company’s feature page? The colors are duller than 90s accounting software, and the only animation is a single pulsing dot…”

(ps: That wasn't me! Defamation! It came up with this on its own!!)

Finally, back to everyone’s most pressing question—

How much does it cost?

Because when you hear “multi-Agent workflow,” your first thought is: this must be expensive. We can’t afford infinite token usage.

Of course, multi-Agent systems consume more tokens than single Agents.

That’s unavoidable—just like choosing HTML over Markdown, better experiences come at a price. Totally normal.

But honestly, the key lies in actual effectiveness.

If it saves time and delivers value, it’s worth it.

And MiniMax is being pretty straightforward this time.

TokenPlan and Agent Plan have been merged.

One subscription unlocks CLI, API, and Agent access—all models including M2.7, music, video, and voice included.

Credits are shared between Agent and API usage—one payment, two functions.

Users who previously subscribed to both plans will receive an additional month of membership for free.

Why Isn't One AI Enough Anymore?

I’m so excited because this solves a pain point I've struggled with for a long time.

If you’re also a vibe coding enthusiast, you’ve definitely experienced these three moments of breakdown:

△ Image generated by AI

Breakdown One: The Agent slacks off.

You ask AI to write a report, and after three paragraphs it stops:

I’ve completed 1/2/3. Should I continue?

As if it doesn’t understand!

Say “continue,” it stops again. Say “continue” again, it stops once more.

By the end of the night, half your time is spent typing “continue,” “continue,” “continue”…

Breakdown Two: Long tasks make the Agent dumber.

At first, it feels like a smart assistant. But as it runs, it turns into someone busy yet easily distracted.

You have to keep reminding—“Do you remember the previous requirement?” “Why did you turn a research task into marketing copy again?”

Breakdown Three: Ghosting…

Send a message to AI via WeChat or Feishu, and either it spits out a shallow answer in 30 seconds, or you stare at the chat window for 10 minutes with zero response.

Hey, why aren’t you replying? What are you doing??

These are high-frequency phrases I often send to my AI colleague (affectionately nicknamed “crayfish”).

Every heavy AI user has lived through these scenarios.

So why are long-running tasks so hard?

MiniMax actually addressed this in their technical blog.

△ Image generated by AI

Simply put, this is a “curse” baked into single-Agent systems from birth.

It mostly comes down to context limitations.

First, single Agents suffer from context anxiety.

This is a deep issue. Training models for ultra-long tasks requires massive investment in money, time, algorithm optimization—resources most teams can’t afford to dedicate.

As a result, models generally lack clarity on when a long task should be considered “complete.”

They fear making mistakes or exhausting tokens, so they stop midway to ask for confirmation.

It’s like working with an overly cautious intern who asks for approval after every tiny step.

Worse, even if you flood it with endless context (as if tokens were free), performance still degrades.

Currently, there’s no solution.

Due to fundamental attention mechanism issues, as context grows longer, the Agent shifts from being helpful to easily distracted.

Context compression becomes necessary—but inevitably loses information and increases user anxiety.

More troubling: single Agents struggle with self-regulation.

They may sincerely self-check, but they’re reviewing their own freshly generated output.

When you’re both player and referee, it’s hard to objectively judge correctness.

Lastly, a very practical problem:

Single Agents can’t respond quickly during long tasks.

You practically can’t run long tasks alongside regular conversations. Once it starts working, it’s hard to communicate via IM.

Long tasks and current chats share the same context. Letting new messages in risks disrupting the ongoing task.

But if you don’t intervene, you’re left waiting.

It’s awkward.

Ultimately, these aren’t model capability issues.

They’re architectural problems.

Which brings us back to Mavis—its Agent Team is designed specifically to solve this architecture problem.

The idea is simple: one main Agent leads, with three roles—Leader, Worker, Verifier—collaborating through分工.

Here’s a key design element: Workers and Verifiers are in adversarial relationship.

The condition for a Worker to stop is precisely what triggers the Verifier to start. The Verifier stops only when it can no longer find issues—and those issues become the reason for the Worker to restart.

Similar to R&D and QA departments in a company, this adversarial iterative process delivers high-quality results.

No need for the CEO (you) to micromanage.

Underneath it all is a state machine called the Team Engine.

When to verify, retry, or stop—are all hard constraints enforced at the engine level, not left to model discretion.

Thus, collaboration isn’t limited to a single function call, but evolves into active push notifications and on-demand queries across multiple rounds of interaction.

Finally, one more design I find incredibly cool:

Agents and humans have equal rights.

Users can perform operations such as prompting, spawning, aborting, and killing on Agents, and Agents themselves also have the ability to do the same to other Agents.

The actual channel for operating an Agent can be a user, another Agent, or the Team Engine.

All actions follow the same protocol. Who did what, and whether there was any privilege escalation, can all be audited and traced.

Of course, for high-risk nodes, human-in-the-loop oversight is still required.

So, once these mechanisms are in place, what kind of outcome can we achieve?

It’s the complete resolution of the three breakdowns mentioned earlier.

1. No more stopping to ask you.

The Leader oversees the overall goal, Workers focus solely on executing subtasks, and termination conditions are controlled by the Team Engine—no longer left to vague model judgments about "is this enough?"

2. No more getting dumber as it runs.

Each Worker has isolated context; a research Agent won’t be polluted by code-writing information. The Verifier reviews from an independent perspective—it's not self-inspection.

3. Instant messaging will never go silent again.

(ps: remember to grant permissions first)

The main Agent instantly acknowledges receipt, then dispatches specific tasks to run in parallel in the background, proactively reporting at key milestones.

You can even add new requirements mid-process:

I just thought of a new direction, blah blah... could you also help me look into this?

The main Agent can immediately reply:

Got it. I'm now launching another group of Agents to investigate. I'll report back as soon as there's progress.

Also, just to update you: 2 out of 5 ongoing tasks are already completed, 2 more are under verification, and 1 is still running.

Honestly, this experience is incredibly hassle-free...

Just like having a Feishu colleague who's always online—no need to ever rush anything.

The Era of Multi-Agent Systems Requires Management

Previously, we were always trying to "raise" an Agent into a superhero—hoping it would become smarter, more capable, able to do everything.

But sometimes I wonder: perhaps an Agent’s capabilities are inherently limited. AI has never been as omniscient and omnipotent as it appears in movies.

If that’s true, then we shouldn’t put so much pressure on a single Agent.

This is exactly what Mavis made me realize most deeply.

Beyond upgrades to the model itself, architectural improvements to Agents can also bring massive enhancements in user experience.

And when we look around, compared to chasing an elusive AGI, what we actually need more urgently is a practical harness tailored to real-world applications.

But this also means that *we*, as the human side of human-AI interaction, must adapt our work habits and ways of thinking accordingly.

You're no longer chatting with an AI.

You’re managing a team.

In the era of multi-Agent systems, everyone needs to learn how to take on a higher-level role.

MiniMax’s design points in this very direction.

In their vision, future Agent products will allow humans to configure Agent roles, capabilities, and boundaries primarily through management dashboards, assigning tasks accordingly.

At that point, the critical skill will no longer be just writing good prompts.

△ Image generated by AI

Finally, let’s stay realistic and talk about "cost-efficiency."

With compute resources still constrained today, every token carries a real price tag. Token consumption versus performance is an unavoidable trade-off.

In fact, MiniMax addresses this directly in their blog—

They don’t hide the fact that multi-Agent systems are “expensive.”

Handoffs cost tokens, sharing costs tokens, aggregation costs tokens… of course they do.

But here’s the problem: when a research Agent gathers dozens of web pages and hands them off to a writing Agent, the information needs to be restructured—

That’s hard.

This isn't something that can be solved simply by making models bigger.

Some problems just require a multi-Agent approach.

So MiniMax’s philosophy has always been practicality first.

Acknowledging cost doesn’t mean giving up—it means using engineering frameworks to manage ROI effectively.

That’s where the Team Engine comes in: deciding when you need an Agent Team, and when a single Agent suffices.

There’s a paper called Cost of Consensus.

It contains a counterintuitive finding: under certain models and homogeneous debate setups, multi-Agent systems may consume 2.1 to 3.4 times more tokens than single-agent self-correction—with no improvement in accuracy.

Without structure, validation, or stopping criteria, a "multi-Agent" setup is just wasting tokens.

That’s not teamwork—that’s an AI chat room.

A team is never the default choice.

For simple tasks, a single Agent is more than sufficient.

Sometimes, even a script is enough.

Not every task requires a meeting.

But when you truly *do* need one, having a reliable team is definitely better than working alone in isolation.

By the way,

MiniMax says they’ll open-source this Agent Team framework, expected to launch alongside MiniMax M3.

Desktop download: agent.minimaxi.com/download

_© All rights reserved. Unauthorized reproduction or use in any form is strictly prohibited._

](https://www.qbitai.com/2026/05/417816.html#)

• [Every lab fears ByteDance, everyone praises DeepSeek! An American researcher’s 36-hour AI journey across China](https://www.qbitai.com/2026/05/414141.html "Every lab fears ByteDance, everyone praises DeepSeek! An American researcher’s 36-hour AI journey across China")_2026-05-08_
• [Markdown is dying… even Karpathy is backing HTML](https://www.qbitai.com/2026/05/416339.html "Markdown is dying… even Karpathy is backing HTML")_2026-05-12_
• [The first batch of “AI-native” undergraduates are about to graduate](https://www.qbitai.com/2026/05/414125.html "The first batch of “AI-native” undergraduates are about to graduate")_2026-05-08_
• [DeepSeek V4’s biggest regret](https://www.qbitai.com/2026/05/412737.html "DeepSeek V4’s biggest regret")_2026-05-03_

Scan to share on Moments

[](https://service.weibo.com/share/share.php?url=https://www.qbitai.com/2026/05/417816.html&title=%E9%87%8D%E7%94%9F%E4%B9%8B%E6%88%91%E5%9C%A8AI%E6%97%B6%E4%BB%A3%E5%BD%93%E8%80%81%E6%9D%BF%EF%BC%9A%E8%AE%A9%E4%B8%80%E7%BE%A4Agent%E4%BA%92%E7%9B%B8PUA&appkey=4017757111&searchPic=true&ralateUid=6105753431 "Share on Weibo")[](https://www.qbitai.com/2026/05/417816.html)

Popular Articles






• About Us
• Join Us
• Contact for Coverage
• Business Cooperation

Scan to follow QbitAI )[](https://weibo.com/qbitai?is_all=1)[](https://www.zhihu.com/org/liang-zi-wei-48/activities)[](https://www.toutiao.com/c/user/53624121633/#mid=1556041376883713)

[](https://www.qbitai.com/2026/05/417816.html#) Tracking new trends in artificial intelligence, reporting breakthroughs in tech industries

QbitAI © 2026 Beijing Geek Partner Technology Co., Ltd. Beijing ICP Record No.17005886-1