How I organized my workspace with neural nets

This may be hard to follow for some, especially those just stepping into the topic. I'll try to explain it as plainly as possible. If you have questions — ask. I try to answer everything. Sometimes questions accidentally get caught in the bin — auto-moderation runs there. I tune it gradually.

Initially I worked the same way most people do. The usual interfaces. ChatGPT. Apps. Chats. But pretty quickly I switched to Claude Code and started working from the terminal. After a couple of months it became clear that even that wasn't enough. One agent isn't enough. Everything starts running too slowly. And the terminal on a home computer still imposes limits. So I decided to move my entire workspace to a remote server.

On the server I spun up a tmux session. The easiest way to picture it is one big window with many small ones open inside. A separate agent runs in each window. The main advantage — they don't shut down. You don't have to relaunch them every time. They're constantly on standby.

The first window is the main agent. You could say it's the system administrator. The whole structure is organized from this window. The system's shared memory — MCP — is also configured here. It works from the root directory and has access to everything. This agent has one main task: to manage the others. It's told what it must always remember, what it can forget, what can be pulled from memory. The most important thing it stores — keys and access credentials. If you imagine all of this as an office building, this agent is the building administrator who has the keys to every office.

Next, the layout of the rooms is built. By the same analogy: the rooms are working directories. A separate project lives in each directory. You can think of them as different companies inside one building.

The next step is deciding who works in each room. For example, my main project is a delivery system. In that room work: a lead coder, several supporting coders, reviewer agents. So a single project may have several agents working at once, each with its own function. When this is set up, tmux automatically creates new windows and launches the right agents.

All agents use shared memory. But not in the same way. Each agent has access only to its own section of memory. It can read other sections, but it can't write to them. This prevents chaos.

Each agent has a short instruction set. That's what it always remembers, even when its context gets compressed. Essentially it's an analogue of human memory: some things we always remember, others we forget. The only difference is that for neural nets, persistent memory is very small. So those instructions need to be written especially carefully.

Once the whole structure was ready, the next question came up. How do I conveniently manage all of this? You can connect from a computer. But that wasn't enough for me. I added access from the phone. Now I can switch between agents and give commands straight from mobile.

The next step — voice. I set up web access and added the ability to give commands by voice. You can send text, voice, images, video. And just explain what needs to be done.

After that I got curious about how to organize interaction between agents. I built a mail system for them. Each agent can send messages to a shared mailbox. And messages automatically end up where they need to. Essentially it's internal communication between the "offices."

But even that wasn't enough. The next step was Telegram. I created a chat, split it into topics and seated the agents there. Now each agent works in its own thread.

There's an important point here. Most AI services offer two payment options: either subscription, or pay per request. Pay-per-request gets expensive fast. So I had to fiddle a bit to get all of this running on a subscription. There are almost no instructions for it, but it can be set up.

So now it looks like this: there are several "companies" and a group of agents works in each one. And I just manage them.

Working with neural nets right now isn't really about knowing technical details anymore. It's more about: imagination, organization, the ability to explain tasks and the ability to control how they're carried out.

And the more I do this, the more I realize one thing. Neural nets aren't a brain. They're a team of interns. And if you don't organize them — they'll create more chaos than benefit.