Member-only story
Most multi-agent tutorials start the same way. #
There’s a planner, a researcher, a writer, a critic, and a coordinator. Each agent gets a role, a prompt, and a neat little box in an architecture diagram. The tutorial then walks through how to wire them together.
What it rarely explains is why there are five agents in the first place.
That’s the more important question.
Every additional agent introduces another LLM call, another prompt to maintain, another context window to manage, another handoff where information can be lost, and another failure point to debug.
A multi-agent system isn’t automatically more capable because it has more agents. Sometimes it’s simply more expensive.
The question isn’t “Can this be split into multiple agents?” It’s “Should it?”
Start with one agent #
Most AI workflows should begin as a single-agent system.
If the task belongs to a single domain, produces a single output, and can comfortably fit into one prompt, adding more agents usually creates complexity without improving results. Consider a document summarization pipeline. You could build a summarizer agent, a quality checker agent, and a formatter agent. Or you could write one well-designed prompt that summarizes, verifies completeness, and produces structured output.