AI Agents#

Course path#

A model produces an answer from its input. To become an agent, it must also be able to observe an environment, choose an action, use a tool, and report the result.

This ability to act raises new questions: which identity should it use? Which permissions should it receive? How can secrets be protected, autonomy limited, actions traced, and the agent stopped when necessary?

We will start from real, bounded experiments to progressively introduce:

• the difference between a model, a script, and an agent;
• tools and authorized actions;
• technical identity and authentication;
• the principle of least privilege;
• state, memory, and traceability;
• safeguards and human validation;
• shutdown, revocation, and rollback.

The first step will be to understand how to give an agent secure and limited access to GitHub. A later step will explore its autonomy on a local system, after the necessary protections have been defined and validated.

Lessons will be added to this plan after they have been experimented with, written, and validated.