Karpathy’s free agent course is two hours; your bookmarks won’t write the loop for you

Andrej Karpathy has released a two-hour free course with a plainly stated topic: agents, harnesses, loops, and graphs. The original post comes from RoundtableSpace, and the video can be played directly. For people doing AI development, this is engineering material you can pause again and again; for nontechnical readers, it can also serve as a conceptual map, so you first know what each of these terms refers to in a system.

Karpathy has long produced AI research and educational content, so his name is familiar to developer readers. General readers only need to know that when he talks about agents, he doesn’t stop at slogans like “let the model do the work itself.” The original post says “everything he knows about agents, harnesses, loops and graphs.” The information in that sentence is concentrated in four terms: an agent is the entity that executes tasks, a harness is the scaffolding around the model, a loop is how a task advances round after round, and a graph is how processes and states are connected. At two hours long, the emphasis is very likely on engineering structure, not on demoing a few prompts.

One reply says: “Two hours of signal. Take notes. Build something. Share results.” It sounds like motivational hype, but the actions are right. With this kind of course, saving and reposting only completes the first step. A more effective approach is to pause and sketch out the agent workflow you currently have: at which step the model is called, how tools are connected, who retries after a failure, and where state is stored. After listening, compare it with your diagram; you’ll find plenty of cases where “you thought the model wasn’t good enough, but the loop simply wasn’t designed well.”

Another reply mentions that agents and loops are the direction of AI, especially for people building trading bots. You can listen to that judgment, but don’t apply it directly. Trading scenarios have requirements for latency, risk control, and backtesting that are far removed from general-purpose agents. The course gives you a sense of structure; it won’t impose domain constraints for you. Another user said they fell asleep at minute 40 and woke up to Karpathy drawing diagrams. Two-hour courses carry this risk, so it’s better to watch in segments and leave one verifiable question for each segment.

In the replies to the original post, some people also ask when the video was first released, and even question “just posted? You mean from 2025?” The material does not provide a definitive answer. Before watching, pay attention to the version, and don’t mistake old conclusions for new progress. Tools and concepts update quickly, and the “everything” in the course title should also be discounted by you.

If you’re building an agent, finish watching first, then change your own loop and harness. If you’re just interested in AI, treat it as a two-hour conceptual tour that will leave you less dizzy from terminology. The video and original post are below; notes can wait, but it’s best to write down questions while you watch.

Original post link: https://x.com/RoundtableSpace/status/2119788347098845694