One day in October, I had about 20 AI agents working at the same time, on four projects. It got chaotic real fast. I couldn’t keep up with which one needed me, or when.
How I got there
Like most people, I started by delegating to one agent. I gave it a task and sat there watching it work. Once I trusted what it could do on a specific kind of task, I tried a second one. Then another, and another, until I found the number I can manage myself. About 20 on simple tasks, or four on something more complex.
At 20, I saw that my MacBook Pro couldn’t keep up with all that processing. So I put an old PC to work as a server, for the heaviest tasks.
I call the agents on that server bees. Each one is a session of Hermes, my agent, which runs on Claude. Giving them their own name makes it easier to delegate from one agent to another. When I tell an agent to hand a task to a bee, it opens a new conversation on the server, named after the task, with the instructions already sent. I see it in my app and can take over like in any other. From October 4 to 19, I opened about 200 conversations on that server, test runs included.
The day I lost track
When you run several sessions on several projects at once, it gets very confusing to know what needs you, and when. A session often ends with something for me to watch for or do later. Once I moved on to another project, I’d forget it, and I had to find the conversation and read it again.
The day of the 20 agents on four projects, I couldn’t keep up. That’s the day I decided to build the For you list.
What crashed
The worst moment was when most of the agents needed a lot of computing resources at the same time. Some crashed, some stalled. I had to slow down and go through each project to get a clear picture of where the work was at.
No real damage. Just a lot of time spent debugging. But I learned a lot of lessons that help me a ton today.
How I keep track now
- Hive Watch. A window that’s always open on my Mac. It shows me every bee and every local agent, and what each one is working on.
- The For you list. The tasks that need me. The bees and my other agents write it, never me. When a bee finishes, it leaves what’s left for me there, in a line or two. When an item is ready for me, my Mac shows a notification.
- Guardrails on every task. So I always stay in control of the scope and of the result I want. Every bee gets the same short brief. The code it writes waits for my review before it goes into the project, and it never opens another bee itself.
- The morning. My client portal holds almost every project and every task left on it. Every morning, I ask my agent what we should work on today. It answers based on deadlines, effort and quick wins. It usually gives me a couple of items it thinks will take me the whole day. In reality they take a few hours, and we easily get through twice what it planned. All without cutting corners on quality, of course.
A lot of people struggle with context switching, jumping from one task to the next. I don’t, really. Mostly because of all my experience playing competitive video games, and how I am as a person. I can keep track of several things at once without feeling overwhelmed or lost.
At first, it was tiring all the same. It’s like any other muscle. The more you train it, the easier it gets, and the more you can do every time.
What I keep for myself
- Clients and secrets. Communication with my clients and sensitive information, like passwords, keys and secrets, stay with me.
- Designing new pages. I don’t let agents design new pages on a website. Most of the time, the result is inconsistent. So first I set the design system, the colours, the type and the components. I build the homepage and we polish it as best we can. Then we make templates that other agents can reuse on the same site.
Where to start
A simple use case for me was a proposal template for new clients, built on the variables I use to price a project. Once it exists, you can put as many agents as you want on incoming leads, one agent per proposal. You can automate it so it runs as each request comes in, or batch them all in one day. Whatever works for you.
I love running several agents at once. It lets me handle my clients’ small requests faster, every day.
