Abstract:

This video is a lesson from the "How to Solve It with Code" series, led by Jeremy Howard and Jonno. It focuses on demystifying and building an AI agent from scratch using the Solvit platform. The core philosophy presented is the "Solvit approach," which emphasizes using AI as a collaborative tool to enhance personal learning and programming skills, rather than as a means to outsource work and degrade one's craft.

The lesson progresses through several stages of building and understanding agents. It begins with a high-level demonstration of a customer service agent that uses custom Python functions as tools to interact with a database. It then shows how an agent with file-system tools can be used to explore and learn an unfamiliar codebase. Finally, the video systematically deconstructs the agentic loop, moving from a high-level library function down to the fundamental, stateless API calls. This reveals the core mechanism: an iterative process of sending conversation history to an LLM, receiving a tool-use request, executing the tool, and sending the result back in an updated history. The final step of creating a complete agent loop from these low-level components is left as a homework exercise for the viewer.

Building an AI Agent from Scratch: A Step-by-Step Guide

• 00:02:21 Course Philosophy and Housekeeping: The instructors answer administrative questions about the Solvit platform (data privacy, billing) and establish the course's goal: to foster skill development through deliberate, project-based practice rather than following a rigid curriculum.
• 00:10:34 The "Solvit Approach": The core philosophy is to use AI to create a positive feedback loop for personal competence. The goal is to leverage AI to learn more deeply and become a better programmer, not to have it do the work for you, which can lead to skill degradation.
• 00:20:07 What is an AI Agent?: An agent is defined as a large language model (LLM) that has access to a set of tools (functions) and can operate in a loop, deciding which tools to use to accomplish a task.
• 00:24:20 High-Level Agent Demo: The first demonstration features a customer service agent.
  • Key Takeaway: By defining simple Python functions (get_customer_info, cancel_order) and making them available to the AI, the agent can perform multi-step tasks like looking up a customer's orders and then iterating through them to execute a cancellation for each one, all from a single natural language prompt.
• 00:30:35 Using an Agent to Learn Code: The instructors demonstrate using an agent with built-in file system tools (rg for searching, view for reading files) to autonomously explore and understand the Claudette library's source code.
• 00:46:15 Creating a Powerful Custom Tool: A custom tool named execute is built, giving the agent the ability to run any arbitrary Python code. This tool is then used to have the agent solve an Advent of Code puzzle.
• 00:58:51 The Danger of Over-Automation: A crucial discussion on the risks of relying too heavily on agents. The instructors argue that outsourcing cognitive work can erode skills and confidence. The recommended approach is to remain "in the loop" and use AI as an assistant to learn and build better abstractions.
• 01:05:53 Deconstructing the Agent Loop: The lesson peels back the layers of abstraction to show how an agent works internally.
  • 01:13:27 Level 1 (High-Level Function): Using the tool_loop function from the Claudette library, which automates the entire agentic process.
  • 01:23:24 Level 2 (Manual Loop): The process is broken down into a manual while loop. This shows the core cycle: the LLM receives a prompt and returns a "tool use" request; the code runs the tool and passes the result back to the LLM, which continues until it generates a final answer.
  • 01:30:06 Level 3 (Raw API Interaction): The final layer uses a low-level client to interact with the API. This demonstrates that LLMs are stateless—they have no memory of their own. The entire conversation history must be manually constructed and sent with every API call.
• 01:49:08 Homework Assignment: Viewers are tasked with building their own do_loop function from scratch, using only the low-level client and manual history management, to solidify their understanding of how agents fundamentally operate.