He Gave 40+ AI Employees 3 Managers; the Comments First Asked: Can He Really Manage It?

Justin shared an AI team architecture: 1 AI general manager, 3 AI project managers, and 40+ AI employees.

The general manager is Claude Opus 5.5. The three managers are GPT 6.1 Sol, Claude Opus 5.5, and Claude Fable 5.1. The employees doing the work use Claude Sonnet 5.5, Gemini 4, Gemini 3.8 Flash, and DeepSeek 4.1 Flash.

The most common question about this division of labor was: what is it running on? Justin replied that the underlying driver is Claude Code, and it calls other models through plugins and the CLI. As for how the general manager and managers decide who does what, he said they first define the scope of responsibilities, and then each level chooses and delegates on its own.

The first serious challenge in the comments came from the account ✪ Tumi API Relay Station: “Your model version numbers have already fast-forwarded to the year after next.” He immediately added that managing 40+ subordinates at the base layer means context synchronization and fault tolerance alone would be enough to give anyone a headache.

The second half of that comment is more worth reading than the version numbers.

Xing Ge AI's judgment was more specific: with a three-layer setup like this, the layer most likely to go wrong is the middle one; once there are more managers, messages start to get distorted in transit. The general-manager layer should assign less work and do more acceptance checks, and the whole thing will be much more stable.

The problem is not that there are too many employees; it's that information gets compressed, rewritten, and embellished as it passes through managers. In human companies this is called middle-management loss, and switching to AI doesn't make it disappear automatically.

nick's concern was more direct: the architecture sounds plausible, but when it actually comes time to work, are those 40+ employees going to save you time, or first produce 40 half-finished deliverables that you have to review one by one? Someone who can manage 3 managers may not be able to manage the hallucinations of employee No. 41.

Someone replied under that comment, saying the management cost is probably all spent on coordination.

There were dissenting voices too. verchena asked why we must use human organizational structures when there is no need for anyone to take the blame. He posted a clip from an interview with Anthropic engineering leader Felix Rieseberg.

What Rieseberg says in the interview lines up with this challenge. He said he doesn't write code now; “I'm just Claude's conveyor belt.” His team took a week and a half to build Claude Cowork. He spends most of his time prototyping, then letting people pick which ideas are worth keeping. He distinguishes automation from augmentation: automation is handing off work you don't want to do; augmentation is letting AI help you do things you couldn't do before—the latter is the part he finds interesting.

He mentioned one criterion: good things have fewer buttons. Whether a product works depends on putting it in front of a few people and seeing whether they can quickly get the hang of it and do things they couldn't do before.

In the context of 40 AI employees, this is actually a bit pointed. In a three-layer management architecture, whether the number of buttons increases or decreases depends on whether you're standing in the general manager's position or the acceptance-check position.

Back to the architecture itself. Its most practical part is what Justin said: define the scope of responsibilities first. AI Work Diary · Xiao F singled out this sentence: if the scope is not defined, the 40+ employees below will easily each do their own thing, and by then even the general manager won't be able to rein them in; the more layers there are, the less you can skimp on this upfront work.

The real cost of this setup lies in writing the boundaries clearly, not in model calls. If you can't write them clearly, 40 employees are just 40 interfaces all talking past each other. If you can write them clearly, the three layers become meaningful.

As for whether it's worth it, the comments have no consensus. Some people are already using herdr plus cc to control other models to do work; some score the models every month and then adjust the next week's selection. Someone else wrapped it up with one line: it's all bullshit; everyone is betting that AI is worse at scheduling than you are.

That's unpleasant to hear, but it happens to point out the implicit premise of this architecture. You have to be better at division of labor than AI before you're qualified to be that general manager.