KakaoPay: Summary of Core Concepts and Architecture Keywords for AI Agents

A KakaoPay FE developer summarized the core concepts and architecture keywords required when introducing AI Agents into development workflows. An LLM is likened to a stateless function that takes input and produces output, while an AI Agent corresponds to the 'body' that receives a goal, formulates a plan, calls tools, and completes tasks.

How Agents Work and the ReAct Pattern

Agents operate based on the ReAct (Reasoning + Action) pattern. This is a structure that acquires information and performs tasks by repeating reasoning and action. While recent models support native function calling (tool calling), the fundamental framework of 'judgment-action-observation-re-judgment' remains the same. Loop termination criteria include the model's own declaration of completion, exceeding the maximum number of iterations or token budget, and passing external validation.

The Importance of Tokens and Context Windows

Understanding Tokens and Context Windows is essential for managing costs and performance. A token is the smallest unit a model uses to process text; Korean text can require several times more tokens than English, so careful cost management is necessary. The context window refers to the total number of tokens included in a single call, such as the system prompt, conversation history, and tool results. Exceeding this limit can result in the loss of previous information or errors.