AI Agent Architecture — ReAct, Memory, Planning and Multi-Agent Systems · Tool-Augmented LLM — Tools, Functions and the Environment

Toolformer and MRKL — How an LLM Learns When and How to Call Tools

Tool-Augmented LLM — Tools, Functions and the Environment

Introduction

Two landmark systems from 2022 established the tool-augmented LLM paradigm: MRKL (Karpas et al., AI21 Labs) presented a router architecture selecting an expert module (calculator, knowledge base, search), and Toolformer (Schick et al., Meta) showed how a model can teach itself WHEN to call an API and how to interpret results through self-supervised data augmentation. This lesson analyses the mechanisms of both approaches, their architectural differences, and their impact on modern function calling systems.