What you’ll get from this lesson: Distinguish three ways to use models, validate your needs first, and only then consider additional equipment.
Prerequisites: No software installation required. Prepare a work example that can be shared publicly or has had sensitive information removed.
I first look at what I want the Agent to do, then choose how to use a model. Subscription-based sign-in, APIs, and local models represent different approaches to usage and management. Separating them first means you do not have to learn all three at once.
For the Subscription Route, First Check Whether the Tool Supports It
The course originally planned to start with a paid ChatGPT account and a specified Pi setup method to reduce the number of things to handle in the first hands-on exercise. Official Codex supports both ChatGPT subscription sign-in and API keys, but that documentation does not, by itself, grant blanket authorization to all third-party tools.[2]
So first confirm the sign-in route against the tools and versions in the course’s current release. Do not casually hand your account credentials to unfamiliar tools. This article only helps you choose a direction. Before actually setting anything up, you still need to review L1’s prerequisites and records of actual testing.
With the API Route, You Manage Usage and Billing
An API is an interface through which programs and services exchange data and call functions. When using a model API, check the provider, model, usage-based billing, authorization method, and restrictions. This route suits people who need to arrange their own program integrations, but it also adds the work of managing bills and credentials.
If you are only trying to understand the concepts for now, you do not need to obtain an API key yet. Once you have chosen a tool, check which access method it supports.
For the Local Route, Consider Both Capabilities and Equipment
Local models run on your own equipment and are affected by factors such as available memory, model version, and how they are run. The sequence I recommended in the original draft was to test functionality with a sample that contains no sensitive information before considering additional hardware.
Even when model names are similar, cloud and local environments may differ. You need to recheck the version, quantization, context length, and tool support. Getting a task to work in the cloud first can reduce blind purchasing, but it does not directly prove that your computer at home can produce the same results.
Evaluate Three Things in One Small Task
Choose a task you will actually use, such as extracting action items from a work note that contains no sensitive information. First check whether the content is correct, then whether it follows your requested format, and finally whether unauthorized actions are blocked.
If you have only tested text quality in a chat window, that test establishes only the text-related part. To confirm that the Agent can operate tools, you must also test tool calls and permissions in the corresponding implementation environment.
Where the Model Runs and Where the Agent Program Runs Are Two Separate Questions
The Agent program can run on your own computer while sending the content it needs to process to a cloud model. Seeing a local folder alone is not enough to conclude that data will not leave your machine. We will map out this distinction again in the privacy lesson.
Start by completing the route comparison card. Write down which route you are choosing for now, why, and what remains unverified. L0 does not require you to learn local deployment all at once. Use one confirmed route to complete your first small task successfully.
👉 Start L1-01: build your first Agent →
L1 is free after you sign in. Start with L1-01, follow it step by step, and you'll have an Agent of your own.
Not ready to start yet? That's fine. Scroll to the bottom and tap "Next lesson →" to keep reading the free articles, then come back to L1-01 when you're ready.
Sources Checked
[2] OpenAI, “Authentication.” Checked on 2026-09-11. The Pi route comes from the course plan and was not set up again for verification in this round.