Like a Project Manager Commanding an AI Team: Manager Mode

Author: 泯泷

Original: https://mp.weixin.qq.com/s/bzZdL23K3C7w1HAbuCdnuA (October 16, 2026)

Previous articles talked about how peer-level AIs compare and critique each other.

But once things get complex, with dozens of subtasks that are intertwined, peer-level coordination breaks down.

At that point you need a role that can make decisions, dispatch work, and catch mistakes.

I call this approach Manager Mode, and it's currently the most common form of multi-AI collaboration.

It's a lot like a project manager

The core is a centralized role called Manager.

Its job is almost identical to a human project manager: first understand the whole task, break it into subtasks, pick people to do them, watch progress, handle surprises, and synthesize everyone's output.

Its cleverest step is an abstraction: register every subordinate AI as a tool the Manager can call.

In its view, calling a subordinate is essentially the same as calling an ordinary tool.

The benefit: to add a new capability, you just build a new AI and register it as a tool; not a single line of the main flow changes. Different subordinates can also use different models and different tools.

The cost: it becomes a single point of failure. All subtasks must flow through its understanding; if the wrong person is chosen or information gets distorted, the whole chain goes off course, and it's easy to lose track of too many things.

It's like a project manager assigning work to specialized subordinates.

The strongest model should go to the one who dispatches

On resource allocation, there's a counterintuitive conclusion here.

Whether a project succeeds often depends less on how fast the workers are and more on whether the work is assigned correctly.

If planning is good enough, even average executors can succeed; if planning is wrong, everyone downstream is doing busywork.

So I recommend keeping the smartest model and the most detailed prompts for the planner, rather than spreading them evenly across every AI.

Two ways to assign work

One is one at a time: finish one task and then decide the next, suited for work with clear sequential dependencies.

The other is to launch several at once and race to see which returns a result first.

One detail: the settlement point should land on the result that has been verified as successful, not the one that merely claims success.

Assigning one at a time: Manager schedules sequentially, subordinates hand results back.

An example I really like: translating a book

Translating a technical book is far more than changing language. You also need consistent terminology across the whole book. If "agent" is translated as "智能体" in one place, later you can't call it "代理".

If you hand this to a single AI, problems appear immediately.

The glossary, translated chapters, its own reasoning, tool outputs keep piling up and soon overflow its small desk.

Worse, it loses its way: a later chapter uses a different translation than the beginning, or it even invents a terminology rule that never existed.

The team-based solution is division of labor plus isolation.

One AI handles terminology only, one handles only the current chapter, and one looks for inconsistencies between earlier and later parts.

The team leader only needs to know the overall plan, progress, and results; it doesn't need to hold every chapter's full text in its head.

Four words: isolation.

Another formation

Besides peer review and leader-assigned work, there's also a decentralized formation: no leader, a group of AIs freely discuss like humans.

For cross-organization collaboration, there's a standard specifically for AI-to-AI dialogue that unifies task states.

How to choose? My rough spectrum is:

When you want to polish a single product, use peer review.

When you want to coordinate a large project, use leader-assigned work.

When you need open collaboration or even cross-organization work, use decentralization.

Closing words

Looking back at this whole journey, it's really one thread.

First ask whether forming a team is worth it, then decide whether to share memory, then pick a formation, and hold the line on independent verification.

Multi-AI is never about more being better.

The real threshold is whether you can think clearly about how information flows between them, how commands travel, and how results are verified.

Reach that level, and you're no longer someone stacking AIs; you're someone designing collaboration.