0:01Before we start building the agent, let's get the code onto your machine.
0:05It's all in one public repository on GitHub. Clone it, and you can run each episode as you follow along.
0:11If you only want to watch, you can skip to the next video.
0:14If you want to run the code, this video covers what you need. We look at the packages, the configuration, how the repository is organized, and how to run an episode.
0:25First, let's look at the packages the code needs.
0:28There are six of them, and there's no agent framework among them, because we write the agent ourselves.
0:34The first is the OpenAI SDK, which our code uses to call the language model.
0:38The same SDK works with any provider that offers an OpenAI-compatible endpoint, and many do. So switching providers only changes the configuration.
0:48python-dotenv reads our settings from a.env file, which we set up in a moment.
0:53tiktoken counts tokens. The agent uses it to record how large each round of tool results is.
0:59The last three are for checking code. The agent runs the project's tests with pytest. Later in the series, it also uses the other two to lint the code and measure test coverage.
1:11There's one configuration step before you can run the code.
1:16This example file lists every setting. You copy it to a file named.env, and fill in your own values.
1:23These two lines choose the provider and the model. The default is GLM 5.2, reached through OpenRouter, and it's the model we use throughout the series. The comments below show the settings for other providers.
1:37Then you fill in the API key for the provider you use, and leave the others empty. With the default settings, that's the OpenRouter key.
1:46The section at the bottom configures the benchmark harness. You only need it for the benchmark part, at the end of the series.
1:53That's all the configuration the code needs. Now let's look at how the repository is organized.
1:59The eval folder at the root is the benchmark harness that the last section of the configuration was for, so we leave it for now.
2:08The agent we build is in the episodes folder.
2:11Each folder in it carries the number of its video, starting with the loop, which is video three.
2:17Each one is self-contained. It holds a complete agent, and its own copy of the project the agent works on.
2:24Let's open the first one.
2:26agent.py holds the agent's code.
2:29We go through it in the next video.
2:31Next to it is the system prompt.
2:35These are the instructions the agent starts every task with.
2:38Beside agent.py and the system prompt is a folder called initial. It holds the project the agent starts from.
2:46It's md2html, the small Markdown-to-HTML tool the agent works on, with its tests.
2:53When you run an episode, the code first copies this folder into a new one called sandbox. The agent does its work there.
3:00So every run starts from the same clean copy, and initial itself never changes. The sandbox folder appears here after your first run.
3:10Let's compare the loop's folder with the next episode's.
3:14The loop's folder has its initial copy of the project, the agent, a README, and the system prompt.
3:21The next episode's folder has the same layout.
3:23It adds one new file, tools.py. Each episode adds the code for what its video teaches, on top of the one before.
3:33Every episode folder has its own agent and its own initial copy of the project. Once you know one, you can find your way around the others.
3:41Next, let's see how to run an episode.
3:45The README has the steps for setting up and running the code.
3:49First, you create a virtual environment, activate it, install the packages, and copy the example.env file we looked at.
3:57Then there are two ways to run an episode. The first is to go into its folder and run its agent.py, which is all you need while you work in one episode.
4:07The second is run.py, at the root of the repository.
4:11You give it an episode's folder, and it runs that episode's agent. It saves the run's tool calls into a logs folder in the episode, with one folder per run, so you can compare runs.
4:23Add the capture option, and it also saves everything the terminal printed during the run.
4:29capture.py, also at the root, is what records the terminal.
4:33In the next video, we start building the agent, beginning with its loop.

Project Setup

Created by Abhyuday DesaiPublished 6 Oct, v2 of 3Runs 4 min 44 sec

0 views

About this video

The setup video for Building AI Agents. We clone the companion repository and walk through what you need before the first episode.

You'll see:

The six packages the code uses, and why there is no agent framework among them
The .env file: choosing a provider and a model, and setting your API key
How the repository is organized: one folder per episode, each a complete agent
The initial folder and the sandbox each run works in
Two ways to run an episode, how to record one, and how to view the page the agent built

The next video starts building the agent with its loop.

What changed

v306 Oct 2026abhydesai5 min 5 sec0480b4e
The video now also covers render.py, which you run after a session to see what the agent built: it takes one of the sample Markdown files from the examples folder and turns it into a web page with the agent's md2html from the sandbox. The code is shown at a newer commit: the initial folder now carries its own README, the README's safety advice is softened, and the source is wrapped to 86 columns. Smaller fixes: rows that appear in the file tree sit under their folder, and visible clicks keep their click sound.

Same chapters. Transcript: 2 of 44 lines changed. 21 sec longer.
Show the differencesHide the differences
1. And render.py is for after a run, to see what the agent built.
  It takes one of the sample Markdown files in the examples folder. Then it turns it into a web page, using the agent's md2html from the sandbox.
0 views0 likes on this version

Watch v3

v206 Oct 2026abhydesai4 min 44 sec5e6c760
The default model is now GLM 5.2, reached through OpenRouter, and the line says plainly that it is the model used throughout the series. The previous cut set DeepSeek V4 Flash as the default, so anyone following along was configuring a different model from the one the series actually runs. The layout scene no longer promises a line-by-line walkthrough of agent.py, only that the next video goes through it. Narration: .env and python-dotenv are spoken "dot ENV" and "python dot ENV" rather than letter by letter, and md2html as "MD to HTML". Visuals: the gitignored rows in the file tree are dimmed less, since at the old setting they were nearly invisible against the light theme.

Same chapters. Transcript: 2 of 44 lines changed.
Show the differencesHide the differences
1. ~~These two lines choose the provider and the model. The default is DeepSeek V4 Flash, reached through OpenRouter, and the comments below show the settings for other providers.~~
  
  These two lines choose the provider and the model. The default is GLM 5.2, reached through OpenRouter, and it's the model we use throughout the series. The comments below show the settings for other providers.
2. ~~We go through it line by line in the next video.~~
  
  We go through it in the next video.
0 views0 likes on this version
v105 Oct 2026abhydesai4 min 42 secf6dc15a

First published version.

0 views0 likes on this version

Watch v1