Lets Build An Agent from Scratch — Kam Lasater

The talk centers on building AI agents without relying on existing frameworks, focusing on understanding the underlying components. It defines an agent as a combination of an LLM, memory, planning capabilities, tools, and a continuous loop. The discussion progresses through stages, starting with basic LLM calls and gradually incorporating tools like SERP AI for Google searches. A key step involves creating a to-do list to manage tasks, enabling the agent to plan and execute actions, such as browsing websites and summarizing content. The agent's programming emphasizes a friendly but direct tone, task-specific focus, and continuous evaluation of goals. The speaker illustrates the agent's capabilities with examples like planning a date, highlighting its potential for interesting applications.