One Agent, Three Agents, or a Workflow? What the Evidence Says Works Better

At equal compute, one strong agent usually beats a small team. The exceptions are real, but they’re narrower than the diagrams on Twitter suggest.

I had three boxes on a whiteboard labelled Researcher, Analyst and Writer, with arrows between them. It looked like a real system. Then a colleague asked one question: “Compared to what?”

I didn’t have an answer. So before writing any orchestration code, I spent a week reading every controlled comparison I could find of single agents, multi-agent teams and fixed workflows. This piece is what I learned, and the test I now run before I pick an architecture.

The architecture I almost built. The arrows looked like progress. They were mostly cost.

Why this question matters now

Multi-agent is the default pitch in 2026. Every framework ships an orchestrator. Every demo has a planner handing work to specialists.

When I checked the numbers, though, most of the headline wins compared a multi-agent system against a single agent that had been given far less compute. Change that one variable and the story changes.

What the controlled studies found