Lately I start my day with just one AI agent. I call it magnus, and I give it the list of everything I want to get done that day. Magnus then launches the agents for each of those items and briefs them. Then it spends the rest of the day managing my attention and bringing me only what needs me.

Magnus is my chief of staff. It started as an experiment in early October on Command Bridge, the tool I built to talk to my agents by voice. Now magnus is the only agent I talk to, and the others keep working without speaking to me. A week in, I told magnus the experiment was no longer an experiment.

The evening fry

Before this experiment I used to keep 3 to 6 agents alive, each in its own lane, and I would speak to all of them. I’d round-robin through them and try to divide my attention equally, so none of them got starved.

At the beginning of the day it was amazing. It worked, I felt productive, and all of the agents were busy and making great progress.

But by the end of the day I was so tired that I would start to favor just one or two agents. The other ones that still needed my feedback were stuck, not making as much progress as I needed them to.

I’d written before that any more than six and my brain fries. A BCG study HBR published gave that a name, brain fry, the mental fatigue of using or overseeing more AI tools than you can handle. In its survey, productivity rose up to three tools and dropped after that.

What the chief of staff does

Magnus keeps me focused and busy. Its job is not only managing all the other agents, but also managing my attention. What it brings me is a draft to approve, a question an agent can’t answer without me, or a result I asked to see. Everything else stays with the agents.

Magnus also keeps a Kanban board of everything I have asked it to do, how each item is moving and what the agents are doing.

Getting magnus to behave the way I needed it to took a lot of training. For example, it needs to bring me one decision at a time and wait for my ruling before moving on. And it can never just give me an update: if the update doesn’t require a decision from me, it has to also feed me something that gets me thinking.

Here are some of the rules that drive Magnus’ behavior:

1. Say what you're about to do before you do it, and keep it terse.
2. Name the thing. Never make me remember what "it" was.
3. Never stop at the finding. Say what it means and what I should do next.
4. Speak when I ask, instruct or hand back. A pause is not an invitation.
5. When an item closes, bring me the next one in the same breath.
6. Keep one board of my asks: waiting on me, being worked, scheduled, done.
7. Every approval you relay quotes my words and the time. Nothing goes out without my OK.

A team roster

A chief of staff alone isn’t enough. I have a team of agents with names, roles and responsibilities, so every agent knows what it has to do and can keep pushing on its own. And magnus already knows which one to launch for each item, based on those roles and responsibilities, and I don’t have to brief an agent again after each goal.

Each role is a project or an area of responsibility in my second brain, with its own notes, guides and startup instructions. The agent is a persistent name that holds the role, so if I put a second agent in the same role, it starts up the same way.

Since I stopped talking directly to every agent, my team went from 6 to 16 named agents. On a busy day I’ve had up to 20 named agents running at once, 100s if we count the short-lived sub-agents they spin up when needed. But not every agent runs every day. Magnus knows what I need to get done, and it starts up or tears down the agents it needs based on what I ask for. When magnus delegates, it hands an agent the outcome, my words and the limits, not the steps. My agents are intelligent and have access to my second brain, so they can reason and ground on their own.

How many leads do you need?

I talk to my chief of staff and, at most, one or two other leads. Most days the chief of staff is enough. Under each one I can have as many agents as needed.

When I need to dig deeper on specifics, I promote an agent to a lead, which means I get to talk to it directly instead of through my chief of staff. This post is an example. I promoted orion, my blog agent, so I could go over the details with it directly. When we’re done, I demote it, and it goes back to working under magnus.

I keep the number of leads small for a reason. In the past I tried a team of agents with a strategist, an architect, a tech lead and a development team. I only spoke with the strategist, and it delegated to the architect, the tech lead and the dev team. In that experiment, the communication between me and the strategist, and between each agent, was so poor that the work was low quality. It lacked my taste. I think it was mostly Opus 5’s fault, because its communication style was so bad. But it also showed me how important it is to never delegate taste.

Since every extreme is bad, a chief of staff is good, but it’s also important to be able to dig deeper by promoting an agent and talking to it directly.

What stays with me

Since my agents have roles and responsibilities, they can keep pushing on their own. But there is a limit to how much they can push without me, and to how much I understand of what they’re doing on my behalf.

Instead of reading walls of text or listening to long explanations of what the agents are trying to explain to me, I have them show me. On a diagram, an animation, a recording of a feature being used, data I need them to render from something they executed, or even a diagram of a concept, and they walk me through it, pointing at what they’re talking about.

That gives me more understanding per minute than reading or listening would. And that understanding is what lets me exert my taste into the work I produce and make sure my intent was fully executed, instead of rubber stamping whatever an agent prepared.