So you've been playing with agents for a while. You started following AI accounts and suddenly you're drowning in discussions about multi-agent systems. You're being told to have:

And before you know it, you're running a company of agents instead of actually saving time.

Look, if you have a company to run, the bandwidth to manage multiple agents, can keep track of updates, shared knowledge, permissions, tools and workflows, go ahead, be my guest.

But if you just want to run agents, experiment, build cool stuff and actually enjoy yourself while doing it, stick to one agent.

The funny thing is that a lot of people building these systems successfully aren't running armies of agents for everything they do.

I fell into the rabbit hole too

I started with one agent. Then I fell into the specialization rabbit hole. I had multiple agents running in parallel. One for research, one for coding, one for planning, one for execution.

Sounds great in theory.

In reality, one agent would learn something and the others wouldn't have access to it. One would have tools another didn't. One would have context the others lacked. I spent more time managing agents than actually using them.

One agent, many routes

Today I've gone back to one agent, many routes. I use Discord channels to keep my head organized around different topics, and the agent decides when to call tools, when to launch sub-agents, when to write code, and when to do research.

If I need it to act like a coder, I ask. If I need it to act like a marketer, I ask. If I need it to act like a business analyst, I ask.

Could I get a slightly better marketer with a dedicated marketing agent? Probably.

Would it be better enough to justify the extra complexity and headaches? Not for me.

My advice

A single agent that's not perfect but actually gets used is better than five supposedly perfect agents that become a nightmare to maintain and get unused in the drawer.

Start with one. Experiment. Build things. Learn what actually limits you. You can always add complexity later.

It's a lot easier to learn from one working agent than it is to untangle five agents that nobody enjoys using.

Learn to drive first. Then buy a sports car. Then worry about Formula 1.

Not the other way around.