He Let One Person Command an Army of Agents, and the Comments Are All Asking Who Reviews the Code

A tweet has been circulating quite a bit on developer timelines recently. The poster says he found the strongest AI engineering workflow ever. The tool is free and open source, called Herdr, and its function can be summed up in one sentence: it lets you command an army of agents.

In the video, he stands in front of a bookshelf with a blue light strip glowing behind him, talking excitedly.

First, let's be clear about the background. Herdr is an open-source agent orchestration tool. According to the tweet, it lets people spin up multiple AI agents to work at the same time, rather than pasting back and forth in one chat window. Someone mentioned that it had just released a desktop app. The poster specifically added: he didn't take money from Herdr, and free open-source tools usually don't pay people to speak for them.

That sentence got called out in the comments. Someone replied: You say the same thing every time—either they paid you or this is the best tool. I write way more code than before. The script needs updating.

The more interesting replies come after that.

One says: Commanding an army of agents is fun, but the real job is being the only code reviewer for that army.

Another one is more direct: Forget the army for a second—I still can't get an agent to center a div. A div is the most basic tag in web layout, and getting an agent to center a div is an old joke in programming circles for "it can't even do the simplest thing."

Someone else asks about technical details: Does each agent get its own branch, or does each open its own git worktree, or do they all edit the same checkout?

These questions all point to the same place. The tool solves how to run a bunch of agents at once. It doesn't solve who merges what those agents hand back, who reviews it, or who judges whether it's right. Spin up ten agents at once and output speed goes up—but the review queue goes up with it. A person sitting at the end of that queue will quickly realize they are the slowest part of the whole pipeline.

Some people in the comments defended it. One user said they had studied Herdr for weeks without trying it, but after watching the video they installed it immediately and it worked great. Another reply moved the scenario outside coding: A small company doesn't need an army; it needs three agents that don't skip work—one answering calls, one handling follow-ups, one watching the data. The common thread is the same sentence: you are no longer the bottleneck.

That sentence is true, but only half true. At the scheduling level, you really are no longer the bottleneck. At the judgment level, you suddenly become the only bottleneck. In writing code, generation has never been the most expensive part. The most expensive part is confirming that what was generated runs, can be maintained, and hasn't quietly changed logic elsewhere.

There's another question that was asked twice and the poster didn't answer: How do these agents avoid stepping on each other's toes? Opening branches and worktrees is the ready-made engineering answer, but every extra branch adds another layer of merging and conflict-resolution cost. Once the number of agents goes up, git itself may become the new battlefield.

At the end of the tweet, he offers a shortcut: throw this video to the Grok bot and it can set it up for you in one step. That's pretty practical, and it also shows the typical onboarding path for this kind of tool now: first set up the environment with natural language, then slowly understand what is actually running inside it.

As for whether it's worth switching, someone in the comments brought up Hermes, Codex, and Claude and asked whether the switching cost and benefits cancel out. There's no uniform answer. Tools are iterating fast; a workflow configured today may have a smoother replacement three months from now. The poster's attitude is clear, though: get it running first and talk later.

I can agree with the "and talk later" part; the "get it running first" part deserves a discount. Commanding an army of agents sounds like the beginning of scale, but in reality it's more like turning the waiting time that used to be serial into waiting time that is parallel. What really determines success or failure is still the person sitting at the end reviewing code—whether he has enough time, context, and patience. The tool raises the ceiling, but it doesn't raise the floor.