Multi-agent systems: when splitting one AI agent into several actually helps

The pitch for multi-agent systems is intuitive: instead of one general-purpose agent trying to do research, write, check facts and format a report, assign each step to a specialized agent and let a supervisor coordinate them, the way a manager delegates to a small team rather than doing everything alone. The idea is not new — researchers at Microsoft formalized it in 2023 in AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation Framework, a framework where customizable agents converse with each other, with tools and with a human to solve a task collaboratively. What the pitch usually leaves out is that coordination is not free.

The cost side of the equation

Every additional agent adds a handoff — a point where one agent’s output becomes another’s input, and where an ambiguity or an error can compound instead of being caught. A single well-scoped agent fails in one place, in one way you can inspect. A chain of five agents can fail in twenty-five, and by the time the output reaches a human, the original mistake may be three steps removed from where it happened. More agents is not automatically more reliable — it is more moving parts.

When splitting genuinely pays off

The cases where it does tend to share the same shape: the subtasks are genuinely different in nature (retrieving information versus writing versus verifying), each one has its own clear success criterion that can be checked independently, and a supervisor — human or automated — can see and interrupt each step rather than only the final output. Splitting a single vague task into three vague agents rarely helps; splitting a well-understood pipeline into stages that were already distinct often does.

• →Is there a genuinely parallelizable or sequential subtask, or is this one job being artificially cut into three?
• →Does each agent have its own success criterion that can be verified independently of the others?
• →Can a supervisor see and interrupt each step — not just review the combined final output?

Before adding a second or third agent to a workflow, the more useful question is usually not "would another agent help here" but "is one well-scoped agent, properly supervised, already enough" — the answer is more often yes than the multi-agent pitch suggests.