Beschrijving
What really separates a plain LLM call from an AI agent?
In this live session, Udacity instructor Peter Kowalchuk answers that by building an agent from the ground up in plain Python — no LangChain, no orchestration library — adding one capability at a time until a single model call becomes a multi-agent workflow.
Starting from one LLM call, Peter layers on persona, knowledge, memory, tools, and finally multiple agents working toward a goal — all using one running example: a relocation advisor helping someone move to a new city. By the end, you'll understand the exact pieces every framework abstracts away, and why that matters when you take agents to production.
Peter Kowalchuk is a Director of Tech Enablement, an adjunct professor of data science and AI at CUNY, and an instructor in Udacity's Agentic AI
Nanodegree program.
💻 Code from this session (free, public):
https://github.com/udacity/agentic-ai-from-scratch-webinar
🎓 Go deeper in the Agentic AI Nanodegree program:
https://www.udacity.com/course/agentic-ai--nd900?utm_source=youtube
⏱ Chapters
00:00 Meet Peter Kowalchuk
03:26 Setting up the LLM: Azure, Key Vault, and secrets
08:33 Structuring the project: services, agents, knowledge, tools, workflows
09:24 Demo 1 — A direct LLM call has no scope
13:27 User vs. system roles, and why memory matters
16:04 Why an LLM call isn't an agent yet
18:04 Demo 2 — Adding a persona (meet Becky, the Austin advisor)
29:53 Guardrails, bias, and hallucinations
32:55 Demo 3 — Building the first agent with memory
37:31 Running the single-agent call
45:10 Demo 4 — The ReAct loop: reason, act, observe
51:10 Multiple agents: collaboration and routing
51:48 One "mega agent" vs. many specialized agents
53:57 Wrap-up
#AIagents #Python #AgenticAI #LLM #MachineLearning