Articles / agentsvideosfree-resources
The best free talks to learn AI agents, in watching order
Twelve free talks and lectures on AI agents in the order I'd watch them, from what an agent is to context, MCP, frameworks and evals.
There are thousands of videos about AI agents on YouTube. Most are framework demos that age in a few months, so picking a few good ones takes longer than watching them.
This list comes from the Zero to AI agent path on GyaanStack Videos, plus four videos from the Deep Agents path. Every pick is from a channel we rate A or S on the site: university courses, the teams that build the tools, and practitioners who publish their evaluation work. The order matters more than any single video. Each one assumes you've seen the ones before it.
The core path is eight videos and about six hours. The deep-agents extension adds about an hour and a half.
Part 1: what an agent is#
Start with definitions. A lot of confusion about agents comes from people using the word for different things.
1. What Are AI Agents? Agentic Systems ExplainedLecture 1 of Stanford's CS329Z. It defines agents properly, which gives you words for everything that follows. Good for anyone who has used ChatGPT but never built a system around a model. The uploader has turned off embedding, so it opens on YouTube.
2. LLMs for Agent Builders: Context EngineeringLecture 2 of the same course. The idea to take away is that what the model sees decides what it does, so building an agent is mostly deciding what goes into the context window. Also opens on YouTube.
3. Building Effective Agents with LangGraphWorkflows versus agents, and when each one fits. Watch this before you build anything, because a fixed workflow is often the right answer and it's much easier to test.
Part 2: tools and MCP#
Once you know what an agent is, the next question is how it reaches the outside world.
4. The Model Context Protocol (MCP)MCP explained by the team that created it. MCP is one standard way to plug a tool into any client that supports it, so you write the tool once.
5. Model Context Protocol (MCP) Explained in 20 MinutesThe same protocol from a practitioner's side. Watch it right after video 4, so you get the same idea from the builders and from a user. If you'd rather read and run code, I wrote MCP from zero with a tested server and client.
Part 3: frameworks and longer-running agents#
6. DSPy vs. LangGraph: AI Agent Frameworks ExplainedLecture 5 of CS329Z compares two frameworks with very different ideas about where the prompt lives. Watch it to pick one for your first project instead of trying both.
7. Context Engineering for AgentsWhat happens to context as an agent runs for many steps, and how to manage it. This picks up where video 2 left off, now with long-running agents in mind.
Part 4: test it before users do#
8. How To Build And Evaluate Search AgentsA search agent built and then evaluated step by step. If you only remember one habit from this list, make it this one: decide how you'll check the agent before you ship it. My written version is Evaluating an AI agent in Python, which runs without an API key.
Extension: deep agents#
"Deep agents" is the name for agents that plan, hand work to sub-agents and keep notes in a file system, so they can stay on a long task. The Deep Agents path on the site has nine videos. Most are short episodes about LangChain's managed product. These four cover the ideas without needing the product:
9. What are Deep Agents?The four things that make an agent "deep" in this sense, which are planning, sub-agents, memory and files.
10. Managed Deep Agents - QuickstartA first deep agent running end to end. Useful to see the moving parts in one place, even if you later build your own.
11. Managed Deep Agents - SkillsSkills are packaged know-how the agent loads only when it needs it, which keeps the context small.
12. Context Engineering for AI Agents with LangChain and ManusHow a team running an agent in production manages its context. Watch this last; it makes more sense once you've built something.
If you only have one evening#
Watch video 3 (31 min), video 4 (19 min) and video 8 (50 min). That's under two hours. You'll know when an agent is the wrong tool, how agents reach tools through MCP, and how to check one before you trust it. Come back for the Stanford lectures when you start building, because they make more sense with a project in front of you.
What this list leaves out#
Three topics aren't covered here on purpose. Multi-agent systems, where several agents talk to each other, are worth learning only after a single agent works and has tests. Memory across sessions depends heavily on the product you build. Agent security, such as prompt injection through tool results, deserves its own list. The GyaanStack agents subject has written material on all three.
How I'd use this list#
Watch one part, then build the smallest thing it allows before moving on:
| After | Build |
|---|---|
| Part 1 | A fixed workflow of two LLM calls, then the same task as a loop with one tool |
| Part 2 | One MCP server with two tools, called from a client |
| Part 3 | The same agent in the framework you picked |
| Part 4 | Ten test questions with checks on which tool was called and with what arguments |
That's roughly a week of evenings. When you finish, the 12 AI engineering projects post has bigger projects to take on next.
The full Zero to AI agent path, with notes on each step, is at videos.gyaanstack.com, and the Deep Agents path is next to it.