Let's build an AI agent
summary
Builds an agent from an echo CLI into a full LLM-driven system that runs tools in a loop, using Gemini 3 Flash Preview, short-term chat memory, and explicit tool declarations (add, date) before wiring in MCP servers (file system, calculator, bash) and progressive-disclosure skills. The core runtime is a compact 12-line while loop that handles model function calls, executes them, and feeds results back; the agent even refactors its own code and still runs. Emphasizes Open RAG, Docling and LangFlow as practical toolsets and the tradeoffs of context cost with many tool declarations.
AI-written summary · about
chapters (19)
- 0:00Welcome
- 1:03AI agents defined
- 2:07Building agent intro
- 6:14Basic text agent
- 10:47Adding short-term memory
- 14:40Introducing tools to interact with the world
- 18:31Agent loop with tools
- 22:39Date tool added
- 25:25MCP tools expansion
- 31:04File system & calculator tools
- 32:53Agent self-rewrite
- 35:00MCP token concerns
- 36:03Agent skills intro
- 38:10Implementing a skill to say things out loud
- 41:51Skills: CLI execution
- 44:38MCP vs agent skills
- 44:59Future of agent harnesses
- 47:50Open source tools
- 49:38Q&A
related talks
49mLet's build an AI agent
32mBuilding an AI Operating System
49mBuilding a coding agent from scratch
01h 48mBuilding Conversational Agents
18mTalking Shit About AI Agents
40mAgents in Action: Advancing Open Source AI Missions
01h 03mAI Coding Agents and how to code them
31mDesigning AI Agents with the Model Context Protocol: From Answers to Actions