From Chatty LLM Agents to Governable Systems: An Architectural Guide

The article is a guide for developers to understand agentic systems: it clarifies the difference between an agent and a regular workflow, and when to use one or multiple agents. It outlines that an agent is defined as model + work loop + tools + state + rules and constraints. The article traces the evolution from the ReAct approach (2022), which introduced the thought-action-observation cycle, to the conversational architecture popularized by AutoGen in 2023, and then to graph-based structures like LangGraph in 2024. It highlights that one agent is often sufficient for many applications, and that multi-agent setups only yield benefits when roles truly differ in context, tools, or permissions. The article also discusses various strategies (ReAct, Plan-and-Solve, ReWOO, Tree of Thoughts, RAP, LATS, Self-Refine, Reflexion, MRKL) and notes that the more predetermined the steps, the less freedom the model needs. It introduces the concept of an 'agent theater' where identical LLM calls pretend to be a team, and emphasizes separating deterministic logic (code) from probabilistic logic (LLM decisions). Finally, it presents a modern agent stack with layers including Agent SDKs, graph runtimes, memory, sandboxing, and observability, and lists examples like OpenAI Agents SDK, Google ADK, LangGraph, Mastra, and others.