Large language model (LLM) agents are AI programs that automate multi-step business procedures using various tools. Engineers have many design choices but little controlled evidence to guide them.
This study compared four designs: a single agent that directly uses tools, one that can load reusable skill definitions, a central coordinator that delegates to sub-agents, and a swarm of specialists that hand off tasks. All were tested on 13 realistic procedures using the same LLM and tools.
No design was universally best; success varied greatly. The simple single agent was often the most efficient. However, a well-built multi-agent system significantly outperformed it (75.6% vs 67.4%), proving implementation quality matters more than architecture. Start simple and improve quality before adding complexity.