Building Applications with AI Agents

Chapter 8. From One Agent to Many

Most use cases start with one agent, but as the number of tools increases, and the range of problems you want your agent to solve increases, introducing a multiagent pattern can improve the overall performance and reliability. Just as we saw that it’s probably not a good idea to put all of your code in a single file, or bundle all of your backend servers into a single monolith, many of the lessons we learned about the principles of software architecture and service design still apply when building systems with AI and foundation models. As you continue to add functionality and capabilities into your agentic system, you’ll soon find the need to break up your system into smaller agents that can be independently validated, tested, integrated, and reused. In this chapter, we’ll discuss how and when to add an agent to your system, and how to organize and manage them.

How Many Agents Do I Need?

Begin with a simple approach, and only add complexity as needed to improve performance. The appropriate number and organization of agents will vary enormously based on the difficulty of the tasks, the number of tools, and the complexity of the environment.

Single-Agent Scenarios

We’ll begin with single-agent systems, which are suitable for tasks that are of modest difficulty, a limited number of tools, and lower-complexity environments. They are also often better when latency is critical, as multiagent systems typically require multiple exchanges between agents, ...