Agent Architectures Explained

Now that you understand the four key capabilities, let's see how they're assembled into working systems. There are three main patterns you'll encounter.

Simple Agent (Single Loop)

The most basic architecture:

1. Receive a task
2. Think about what to do
3. Take an action
4. Check if done
5. If not, repeat from step 2

Best for: Straightforward tasks that don't require extensive coordination. Like a personal assistant that handles one request at a time.

Limitation: Can get stuck in loops or make errors without catching them.

Reflection-Enhanced Agent

Adds a self-review step:

1. Receive a task
2. Think and plan
3. Take an action
4. Review the output critically
5. Improve if needed
6. Continue or complete

Best for: Tasks requiring quality output, like writing, analysis, or code generation where accuracy matters.

Example: A report-writing agent that drafts, then reviews for gaps, then revises before presenting the final version.

Multi-Agent System

Multiple specialized agents coordinating:

• Orchestrator Agent: Manages the overall task, delegates to specialists
• Specialist Agents: Each handles a specific domain (research, writing, analysis)
• Reviewer Agent: Quality checks outputs before they're finalized

Best for: Complex workflows that benefit from specialization. Like a newsroom where different reporters handle different beats.

Example: Anthropic's research on long-running agents uses a two-agent pattern—one agent works on the task while another monitors progress and handles errors.

How to Recognize These in Products

When evaluating AI tools, ask:

• Does it just respond, or does it take multiple steps?
• Does it review its own work?
• Are there specialized components working together?

Understanding the architecture helps you predict what a tool can handle and where it might struggle.

In the next module, we'll dive deeper into how agents actually think and make decisions.

:::