🦹 Agent architectures

AI agents can consist of a single LLM as the decision maker or multiple LLMs (or same LLM, different prompts) working together to solve complex problems.

Let's look at a few different agent architectures:

Tool-calling

This is the architecture most agents start with. It consists of a single LLM that has access to several tools to perform a range of tasks.

If you start with tool-calling agents, but upon thorough evaluation find that you need a more sophisticated architecture, only then consider multi-agent architectures. Bear in mind that fully autonomous multi-agent workflows mean higher costs, latency, and a system that is hard to debug, so use them with caution.

Supervisor

In this architecture, a single agent (supervisor) interfaces with a group of agents to determine the next course of action.

Network

In this architecture, each agent can communicate with every other agent and decide which one to call next or end the execution.

Custom

In this setup, you can decide which agents can interact with each other and how the control flows between them.