How AI Agents Actually Work — React Loop, Tools & Guardrails

Most explanations of AI agents are theory. This is code. In this session I walk a developer friend through how AI agents actually work; from the ReAct loop to tools, guardrails, feedback loops, and a real open source agent codebase. No slides. No hype. Just a real conversation with real questions. What's covered: → Model vs agent — what's actually the difference → The ReAct loop in plain English and pseudo-code → How GitHub Copilot works under the hood → What a "tool" actually is (the JSON definition) → When to use a workflow vs an agent → Guardrails — what your agent should never be allowed to do → Feedback loops — why they make or break an agent → LLM as a judge → Code walkthrough inside a real open source agent If you're a developer getting started with agents, or a founder evaluating AI for your business — this is where to start. —— 🔗 Working on an AI problem? → https://nond.ai —— CHAPTERS: 0:00 - Intro 0:30 - Context: ChatGPT to DeepSeek — how agents emerged 1:40 - Model vs Agent: what's the difference 3:30 - The ReAct Loop explained 5:30 - How GitHub Copilot actually works under the hood 7:30 - What is a "tool" — the JSON definition 11:00 - Model vs Agent: the brain/body analogy 12:00 - Where to start: use existing agents first 13:30 - Guardrails: what your agent should never do 17:30 - Agents vs design patterns — why determinism matters 18:30 - When to use a workflow vs an agent 22:00 - Real example: KYC name matching 23:30 - Hybrid approach: workflow + LLM in one step 27:00 - Feedback loops: why they make or break an agent 28:30 - Coding agent feedback: builds, tests, browsers 32:30 - LLM as a judge 35:00 - Code walkthrough: inside a real open source agent 40:00 - Tool parsing and message history in code 43:00 - System prompts explained 45:00 - Open source agents to explore next —— Repositories to explore: https://github.com/earendil-works/pi https://github.com/anomalyco/opencode TAGS: AI agents, agentic AI, LLM agents, how AI agents work, ReAct loop, AI tools, AI guardrails, coding agents, GitHub Copilot explained, Claude Code, AI for developers, agentic framework, workflow vs agent, LLM as a judge, AI engineering, Nond.ai, Gaurav Gat