Why Many Agents?

Juno keeps growing. Forty tools. A prompt that's ten pages long. And it's started to get confused, using the wrong tools, and mixing up rules. Sound familiar? A startup with one person doing every job has the same problem. The answer is the same as well. Specialise.

• Forty tools, ten pages of prompt
• It's getting confused
• Specialise

Why split an agent up

Splitting helps for four reasons. Each agent has a focused prompt, and only a few tools, so it makes better choices. Each one has a clean context, which isn't cluttered by other people's work. Each one can use a different model. And you can test each one separately. That's the upside.

The real cost

But be honest about the costs. Agents have to pass information to each other, and things get lost, just as in a company. There are more model calls, so it's slower and it costs more. Failures are harder to trace. Most teams that jump to many agents would have been better off with one good agent, and better tools.

When to go multi-agent

So when should you do it? Split when one agent has too many tools to handle, when jobs need clearly different skills, or when tasks can run in parallel. Keep one agent if a single prompt can handle it, or if the steps depend heavily on each other.

Summary