Agent Architecture

Introduction

A production agent is more than a while-loop. The canonical architecture has a planner, executor, memory, tool layer and observer. Each is independently testable and replaceable.

Beginner analogy: a kitchen: head chef plans, line cooks execute, prep station = memory, pantry = tools, expediter = observer.

Understanding the topic

Core concepts:

• Planner: breaks the goal into steps.
• Executor: runs one step at a time.
• Memory: short-term scratchpad + long-term store.
• Tool layer: validates and calls tools.
• Observer: logs, evals, cost tracking.

Syntax reference

Visual workflow / architecture:

┌──────────┐
│  Planner │──► step list
└────┬─────┘
     ▼
┌──────────┐  call   ┌────────┐
│ Executor │────────►│ Tools  │
└────┬─────┘         └────────┘
     ▼
┌──────────┐
│ Memory   │ ◄──── observe + persist
└────┬─────┘
     ▼
┌──────────┐
│ Observer │ logs · evals · cost
└──────────┘

Real-world use

LangGraph, CrewAI, AutoGen, LlamaIndex Workflows, Pydantic AI — all expose these layers as composable pieces.

Best practices

• Build the loop first as plain code; framework later.
• Keep the planner small + critique it with a second LLM.

Common mistakes

• Coupling memory and execution — makes debugging impossible.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Name the five layers of agent architecture.
• Q2. Why separate planner and executor?
• Q3. Scenario: your executor occasionally hangs. Where do you add timeouts?