Building AI Agents that actually work (Full Course)

1mo ago

I sit down with Remy Gaskell to break down how anyone can build AI agents to run entire departments of their business. Remy walks through the core concepts: agent loops, context files, memory, MCP tool connections, and skills. We put everything together by building a fully functional executive assistant live on screen. This is a beginner-friendly crash course that covers Claude Code, Codex, Cowork, Antigravity, Manus, and OpenClaw, showing that once you understand how to "drive," you can jump into any agent platform. By the end, listeners know exactly how to set up markdown-based context files, connect their everyday tools, and create reusable skills that compound over weeks and months. Timestamps 00:00 – Intro 01:35 – Agents vs Chat 03:22 – The Agent Loop 05:46 – How Agents work 06:39 – Demoing Agents (Claude Code, Codex, Antigravity) 08:52 – Security and Agent Permissions 10:43 – Comparing Results Across Three Platforms 13:57 – Startup Idea: Cold Email Website Offer 14:50 – Folder Structure and Department-Based Agents 15:52 – Onboarding an Agent Like a Real Employee 17:05 – Voice-to-Text With Monologue and WhisperFlow 18:04 – Chat Memory vs. Agent Memory 19:34 – Building the agents md 22:20 – Context Engineering Over Prompt Engineering 24:29 – How Memory Compounds and Reduces Errors 30:27 – How Big Can memory md Get? 31:43 – Connecting Tools via MCP (Model Context Protocol) 34:49 – Working in Claude Code for High-Value Tasks 37:09 – Why the Real Value Is in Stacking, Not Summarizing 40:04 – What Are Skills? (SOPs for AI) 43:08 – Creating Skills 48:36 – Real-World Example: Ads Analyst Skill: 4-Hour Process in Minutes 50:37 – Chaining Skills together 52:01 – Real-World Example: Automated Car Search 53:34 – OpenClaw and Migrating Agents to More Autonomous Platforms 55:19 – Which Platform Should Beginners Start With? 56:28 – Global vs. Project-Level Skills, Context, and MCPs Key Points * Agent platforms (Claude Code, Codex, Cowork, Antigravity, Manus, OpenClaw) are all running the same observe-think-act loop under the hood — learning one means you can use any of them. * The shift from chat to agents requires moving from prompt engineering to context engineering: load the agent with rich context so simple prompts produce excellent results. * A memory md file creates a self-improving loop where the agent learns preferences across sessions and makes fewer errors over time. * MCP (Model Context Protocol), built by Anthropic, acts as a universal translator between your agent and every tool it needs — Gmail, Calendar, Stripe, Notion, and more. * Skills are reusable SOPs packaged as markdown files; once you explain a process once, you can invoke it repeatedly, and they compound as you add three to five per week. * Scheduled tasks turn skills into automated workflows — morning briefs, car searches, ad library analyses — that run on a cron without any manual trigger. Numbered Section Summaries 1. The Agent Loop in Action Remy kicks off with a live demo, sending the same prompt — "build a minimalist portfolio site for Greg Isenberg" — to Claude Code, Codex, and Antigravity simultaneously. All three platforms run the same observe-think-act loop: research the subject, write the code, spin up a preview, and verify the result with a screenshot. The demo makes it tangible that every agent harness is just a different car with the same engine. 2. Onboarding Your Agent Like a Real Employee Remy shows that without context, an agent asked to "write me a cold email" has no idea who you are or what you sell. The fix is an agents.md (or Claude.md) file — a persistent context document loaded at the start of every session. You fill it with your role, business details, tools, and working preferences, and the result is that a two-word prompt produces a fully informed output. 3. Memory That Compounds Chat models store memory invisibly in the cloud; agents require you to build it intentionally. Remy adds a memory.md file and a simple instruction i