0:00In this series, we build an AI agent from scratch, in plain Python. First, let's look at what an agent does.
0:07Every agent run starts with a task, such as adding a feature to a codebase.
0:12The task goes to a language model. On its own, a model only reads text and writes text.
0:18So when the model replies, our program checks one thing. Did it ask to use a tool?
0:23If it did, our code runs that tool. A tool is a function, such as one that runs a shell command.
0:30The tool's output is added to the conversation as a new message.
0:34Then the whole conversation goes back to the model, which decides what to do next.
0:39That's one round, and it repeats as the model works through the task.
0:44When the model replies without asking for a tool, the loop ends, and that reply is the agent's answer.
0:51Everything else in the series builds on this loop. It's the same loop behind agents in many different fields, so let's look at a few.
0:59A research agent searches, reads what it finds, and searches again until it can answer. Browser and data agents work step by step in the same way.
1:08So can a support agent, with access to a store's order system.
1:12In this series, we build a coding agent, because we can check its work by running tests.
1:18What you learn carries over to the others.
1:20Here's how the series builds it.
1:22There are five episodes, each adding one capability.
1:26The first builds that loop in code, with a model and a single tool that runs shell commands.
1:32The second gives the agent proper tools for working with files, and a short way to define new ones.
1:38The third adds skills. A skill is a set of instructions and tools that the agent loads only when a task needs it.
1:45The fourth adds subagents. The agent can hand part of a task to a separate worker, and several workers can run at once.
1:52The fifth adds checks the work must pass before the agent may stop, instead of trusting the model's word.
1:58Every episode works on the same codebase.
2:02It's a small command-line tool called md2html.
2:05You give it a Markdown file, and it writes out the same document as HTML.
2:10Inside, it does that in three stages.
2:12A document passes through each of them in turn.
2:15First, it splits the Markdown text into tokens.
2:19Then it builds a tree of the document's structure from those tokens.
2:23And it turns that tree into HTML.
2:25It's small enough to follow, and every stage has its own tests.
2:29The agent's tasks are real features in this codebase, and the tests tell us whether the work is right.
2:35Once the agent is finished, we test it on code it has never seen.
2:39For that, we use a public benchmark for coding agents.
2:42Its tasks come from well-known Python projects, and Django supplies the largest share.
2:48Each task is a real issue that someone reported on one of them.
2:52A fix counts only when the issue's tests pass and nothing that worked before breaks.
2:57That run comes at the end of the series.
2:60By then, you'll be able to do four things.
3:02First, you'll write a working agent yourself, starting from an empty file.
3:06You'll know how to add to it deliberately, and why each addition is there.
3:11When it goes wrong, you'll know where to look, because every run keeps a record of what the agent did.
3:16And when you meet an agent framework or product, you'll recognize the loop inside it.
3:22The code for all of it is public.
3:24It's in this GitHub repository, and you can clone it and run every example yourself.
3:29Each episode has its own folder, holding the complete agent as it stands at that point in the series.
3:35In the next video, we set up the project and look at how the repository is organized. After that, we start building.

Series Overview

Created by Abhyuday DesaiPublished 6 Oct, v2 of 3Runs 3 min 49 sec

0 views

About this video

The overview of Building AI Agents. We start with what an agent does: a task goes to a language model, our program checks whether the model asked to use a tool, runs it, adds the result to the conversation and goes around again, until the model replies without asking for a tool. That loop is the same in research, browser, data and support agents. This series builds a coding agent, because its work can be checked by running tests.

You'll see:

The agent loop, step by step: task, model, tool call, result, and the final answer
Agents people use today, and the loop each one runs
The five episodes, each adding one capability: the loop, tools, skills, subagents, and checks the work must pass before the agent may stop
md2html, the small Markdown-to-HTML tool the agent works on throughout, and its three stages
The benchmark run at the end: real issues from well-known Python projects, on code the agent has never seen
What you can do by the end of the series, and where the code is

No code yet. The next video sets up the project and the repository.

What changed

v306 Oct 2026abhydesai3 min 50 secbe9e224

The labels on the card showing where the same loop turns up now stack instead of crowding onto one line. The narration is unchanged.

Same chapters and transcript.

0 views0 likes on this version

Watch v3

v206 Oct 2026abhydesai3 min 49 sec5e6c760

The close now lifts the repository out of the frame and puts "Next / Project setup" in its place, so the video ends by naming what follows rather than on the repo, and the last line holds a second longer. The two flow diagrams, the agent loop and the md2html pipeline, are rebuilt: cards fitted to their labels, the whole flow centered, and the pulse that travels along it runs at an even speed instead of varying. Narration: md2html is spoken "MD to HTML" rather than spelled out letter by letter.

Same chapters and transcript.

0 views0 likes on this version
v105 Oct 2026abhydesai3 min 49 sec19a8bc2

First published version.

0 views0 likes on this version

Watch v1