Arize Agent Mastry Course: Agent Architectures

In this lab, we’ll explore agent architectures by implementing two common frameworks. Understanding different architectures is key to identifying which approach best fits your workflow and use case.

We’ll use the same setup as before, then leverage features of the Agno framework to demonstrate both the Orchestrator–Worker architecture and a Parallelization architecture. Finally, we’ll examine the traces of each framework within Arize to better understand their behavior and performance.

Set Up

Define Tools

Agent Architecture 1: Orchestrator-Worker Framework

In the Orchestrator–Worker framework, we will structure our system using multiple specialized sub-agents (one for each of our 3 tools).

Each sub-agent focuses on a specific capability, such as getting essential information, estimating budgets, or suggesting local experiences.

A centralized orchestrator agent coordinates these sub-agents by delegating tasks to the appropriate one and then synthesizing their outputs into a cohesive final response. This approach mirrors how complex workflows can be broken down into smaller, focused tasks that work together seamlessly.

Tracing this agent reveals the Orchestrator–Worker workflow in action. We can observe how tasks are explicitly delegated to individual sub-agents and how their outputs are combined to produce the final response.

Agent Architecture 2: Parallelization Framework

In the Parallelization framework, we run all sub-agents concurrently instead of sequentially. Each sub-agent works independently on its assigned task — for example, retrieving essential information, estimating budgets, or finding local experiences — while the main agent waits to gather their results. Once all sub-agents complete their work, the agent synthesizes their outputs into a unified response.

This approach offers a significant latency advantage, as parallel execution reduces overall response time without compromising the quality or completeness of the final answer.